diff --git a/dev/trainer_rank_cost_calibration_certificate_attn-moe-h2048-h200-bf16.json b/dev/trainer_rank_cost_calibration_certificate_attn-moe-h2048-h200-bf16.json new file mode 100644 index 000000000..d5af4a8e5 --- /dev/null +++ b/dev/trainer_rank_cost_calibration_certificate_attn-moe-h2048-h200-bf16.json @@ -0,0 +1,140 @@ +{ + "admitted": {"device_classes":[{"capability":[9,0],"memory_class":"hbm-141g"}],"geometries":[{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4}],"param_dtypes":["torch.bfloat16"],"shapes":[[1,1,1,1],[1,2,1,1],[1,2,2,1],[1,4,1,1],[1,4,2,1],[1,4,4,1],[2,1,1,1],[2,1,2,1]]}, + "cells": [ + {"candidates":[{"cp_plan":{"max_rank_tokens":12204,"summary_ms":0.0005159527063369751,"wave_count":0},"fallback_work":602166784,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1538.4658813476562,"n":8,"spread_pct":12.999852219045097},{"cp_plan":{"max_rank_tokens":23926,"summary_ms":0.00035390257835388184,"wave_count":0},"fallback_work":1179826944,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":2643.4248046875,"n":8,"spread_pct":3.284175025961748},{"cp_plan":{"max_rank_tokens":35778,"summary_ms":0.00033527612686157227,"wave_count":0},"fallback_work":1763893504,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":3840.4708251953125,"n":8,"spread_pct":1.5597592674416059},{"cp_plan":{"max_rank_tokens":12154,"summary_ms":0.00035390257835388184,"wave_count":0},"fallback_work":599932160,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1585.5200805664062,"n":8,"spread_pct":19.31614467031251},{"cp_plan":{"max_rank_tokens":47640,"summary_ms":0.0010076910257339478,"wave_count":0},"fallback_work":2348223488,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5060.262451171875,"n":8,"spread_pct":25.382592180619955},{"cp_plan":{"max_rank_tokens":12114,"summary_ms":0.000648200511932373,"wave_count":0},"fallback_work":597960960,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1537.9014892578125,"n":8,"spread_pct":11.607622152148563},{"cp_plan":{"max_rank_tokens":12109,"summary_ms":0.00039301812648773193,"wave_count":0},"fallback_work":597943936,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1571.0377807617188,"n":8,"spread_pct":75.42098773331642}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11920,"summary_ms":0.000391155481338501,"wave_count":0},"fallback_work":587909120,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1526.442138671875,"n":8,"spread_pct":147.1925322924802},{"cp_plan":{"max_rank_tokens":23728,"summary_ms":0.0009927898645401,"wave_count":0},"fallback_work":1169577984,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2578.8602294921875,"n":8,"spread_pct":3.105638955330116}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12138,"summary_ms":0.0006910413503646851,"wave_count":0},"fallback_work":598291712,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1423.1102294921875,"n":8,"spread_pct":1.2647923734523554}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":20701,"summary_ms":0.0003427267074584961,"wave_count":0},"fallback_work":1021030016,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":2396.8580322265625,"n":8,"spread_pct":5.106548943239851},{"cp_plan":{"max_rank_tokens":16154,"summary_ms":0.0002980232238769531,"wave_count":0},"fallback_work":797183232,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":1900.2673950195312,"n":8,"spread_pct":46.567273231136106},{"cp_plan":{"max_rank_tokens":14771,"summary_ms":0.0002253800630569458,"wave_count":0},"fallback_work":729258368,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1801.4177856445312,"n":8,"spread_pct":17.90511346162245},{"cp_plan":{"max_rank_tokens":21963,"summary_ms":0.00019744038581848145,"wave_count":0},"fallback_work":1083450752,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":2348.4730224609375,"n":8,"spread_pct":7.393278472012467},{"cp_plan":{"max_rank_tokens":27497,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":1356166272,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":2787.5731201171875,"n":8,"spread_pct":3.5830670956597572},{"cp_plan":{"max_rank_tokens":35243,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":1737659776,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":3503.2650146484375,"n":8,"spread_pct":2.1125451923040406},{"cp_plan":{"max_rank_tokens":46153,"summary_ms":0.0008326023817062378,"wave_count":0},"fallback_work":2275075200,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":4615.808837890625,"n":8,"spread_pct":1.1399871778549782},{"cp_plan":{"max_rank_tokens":15600,"summary_ms":0.00019744038581848145,"wave_count":0},"fallback_work":770111488,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1865.4463500976562,"n":8,"spread_pct":285.9416205743372},{"cp_plan":{"max_rank_tokens":14494,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":715837184,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1745.09375,"n":8,"spread_pct":6.655209695037874}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16800,"summary_ms":0.00033527612686157227,"wave_count":0},"fallback_work":828526592,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":1957.8742065429688,"n":8,"spread_pct":39.02621036080049},{"cp_plan":{"max_rank_tokens":18971,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":935513472,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":2002.5237426757812,"n":8,"spread_pct":0.5499225236075241},{"cp_plan":{"max_rank_tokens":30232,"summary_ms":0.000867992639541626,"wave_count":0},"fallback_work":1490226176,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":3106.0118408203125,"n":8,"spread_pct":0.5989439975689786},{"cp_plan":{"max_rank_tokens":12255,"summary_ms":0.00025331974029541016,"wave_count":0},"fallback_work":604778368,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":1475.9418334960938,"n":8,"spread_pct":0.792016121525289}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13078,"summary_ms":0.00036694109439849854,"wave_count":0},"fallback_work":645368576,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1681.8115234375,"n":8,"spread_pct":7.4724839229462745},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.00027567148208618164,"wave_count":0},"fallback_work":1233476096,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":2764.640869140625,"n":8,"spread_pct":2.005171508328926},{"cp_plan":{"max_rank_tokens":36716,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":1810249216,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":3954.223388671875,"n":8,"spread_pct":0.8849505068965831},{"cp_plan":{"max_rank_tokens":48890,"summary_ms":0.000260770320892334,"wave_count":0},"fallback_work":2410183936,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":5267.0068359375,"n":8,"spread_pct":13.716680433454501},{"cp_plan":{"max_rank_tokens":12899,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":636776832,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1596.505615234375,"n":8,"spread_pct":8.107833468694675},{"cp_plan":{"max_rank_tokens":12722,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":628054272,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1570.4830932617188,"n":8,"spread_pct":11.243228201849185},{"cp_plan":{"max_rank_tokens":61066,"summary_ms":0.0007171183824539185,"wave_count":0},"fallback_work":3009987840,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":6556.490966796875,"n":8,"spread_pct":1.4784072931065935},{"cp_plan":{"max_rank_tokens":13015,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":642493312,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1669.9314575195312,"n":8,"spread_pct":6.855677034309416},{"cp_plan":{"max_rank_tokens":12701,"summary_ms":0.00027567148208618164,"wave_count":0},"fallback_work":627248768,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1625.0217895507812,"n":8,"spread_pct":8.818068357481836},{"cp_plan":{"max_rank_tokens":12700,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":627428864,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1602.9439697265625,"n":8,"spread_pct":7.9775928035245816}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.3105979934334755,"wave_count":2},"fallback_work":601777600,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1320.8770751953125,"n":16,"spread_pct":2.067551550857737},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":12.868216959759593,"wave_count":1},"fallback_work":1178875104,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":1710.3878784179688,"n":16,"spread_pct":40.00296899028816},{"cp_plan":{"max_rank_tokens":20480,"summary_ms":5.659256828948855,"wave_count":1},"fallback_work":1762372768,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":2338.2376708984375,"n":16,"spread_pct":0.5236482491660203},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":7.145793177187443,"wave_count":2},"fallback_work":599610912,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1311.9741821289062,"n":16,"spread_pct":50.22081716610332},{"cp_plan":{"max_rank_tokens":27136,"summary_ms":10.496336035430431,"wave_count":1},"fallback_work":2346067840,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3076.3233642578125,"n":16,"spread_pct":0.7267340525016208},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.574079928919673,"wave_count":2},"fallback_work":597641632,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1309.9478149414062,"n":16,"spread_pct":1.9857530815245512},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.5770049802958965,"wave_count":2},"fallback_work":597690464,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1308.116455078125,"n":16,"spread_pct":2.118701877270534}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.832671817392111,"wave_count":2},"fallback_work":587468032,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1286.0106201171875,"n":16,"spread_pct":54.93236401825952},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":10.013560997322202,"wave_count":1},"fallback_work":1168504576,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1694.4683837890625,"n":16,"spread_pct":1.341400906341255}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.869758944958448,"wave_count":2},"fallback_work":597741856,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1275.4826049804688,"n":16,"spread_pct":3.4246860163750634}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13824,"summary_ms":9.329071966931224,"wave_count":1},"fallback_work":1020265824,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":1661.5503540039062,"n":16,"spread_pct":1.6003472965955317},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.4620659686625,"wave_count":2},"fallback_work":796702752,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":1457.2014770507812,"n":16,"spread_pct":43.176967972366164},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":5.917434114962816,"wave_count":2},"fallback_work":728909888,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1396.6144409179688,"n":16,"spread_pct":51.15562881796281},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.3302170019596815,"wave_count":1},"fallback_work":1082691520,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":1527.233642578125,"n":16,"spread_pct":1.7481998782201211},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":8.016564883291721,"wave_count":4},"fallback_work":1355141408,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":1796.590576171875,"n":16,"spread_pct":1.6706720561067072},{"cp_plan":{"max_rank_tokens":17835,"summary_ms":5.715635139495134,"wave_count":1},"fallback_work":1736197568,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":2144.8082275390625,"n":16,"spread_pct":0.5214035140839335},{"cp_plan":{"max_rank_tokens":24649,"summary_ms":8.103383937850595,"wave_count":1},"fallback_work":2273023776,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":2817.19482421875,"n":16,"spread_pct":1.7315973422836706},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":7.1755279786884785,"wave_count":3},"fallback_work":769723136,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1456.5579833984375,"n":16,"spread_pct":2.760112345472805},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":7.351061096414924,"wave_count":3},"fallback_work":715567456,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1423.0475463867188,"n":16,"spread_pct":46.401287125387796}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11776,"summary_ms":6.492273882031441,"wave_count":2},"fallback_work":827884032,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":1485.9249267578125,"n":16,"spread_pct":1.6536926861812136},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":6.4023497980087996,"wave_count":1},"fallback_work":934766784,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":1434.4313354492188,"n":16,"spread_pct":2.596430908491263},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":8.88251606374979,"wave_count":1},"fallback_work":1488873344,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":1965.572998046875,"n":16,"spread_pct":32.305150908826455},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.149525986984372,"wave_count":2},"fallback_work":604419584,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":1307.3157958984375,"n":16,"spread_pct":50.62284240084837}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10240,"summary_ms":6.301890127360821,"wave_count":2},"fallback_work":644970208,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1395.3900756835938,"n":16,"spread_pct":56.32784536449038},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":8.10044095851481,"wave_count":1},"fallback_work":1232504896,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":1766.87353515625,"n":16,"spread_pct":1.1755307433322177},{"cp_plan":{"max_rank_tokens":21504,"summary_ms":6.399387028068304,"wave_count":1},"fallback_work":1808716224,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":2450.1280517578125,"n":16,"spread_pct":1.1520543434607038},{"cp_plan":{"max_rank_tokens":25850,"summary_ms":8.334027137607336,"wave_count":1},"fallback_work":2408066592,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":3248.7518310546875,"n":16,"spread_pct":0.7449751115728603},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.7276650331914425,"wave_count":2},"fallback_work":636452672,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1371.2822265625,"n":16,"spread_pct":51.87708349208406},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.666621029376984,"wave_count":2},"fallback_work":627738528,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1361.4514770507812,"n":16,"spread_pct":52.24017318014719},{"cp_plan":{"max_rank_tokens":33280,"summary_ms":8.05666297674179,"wave_count":1},"fallback_work":3007220512,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3869.537109375,"n":16,"spread_pct":0.7846558378924321},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.863664206117392,"wave_count":2},"fallback_work":642163584,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1385.1300659179688,"n":16,"spread_pct":53.63005883809405},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.738111911341548,"wave_count":2},"fallback_work":626999648,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1362.0724487304688,"n":16,"spread_pct":2.2158347334749493},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.776841193437576,"wave_count":2},"fallback_work":627245248,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1360.9933471679688,"n":16,"spread_pct":2.865554051469347}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7680,"summary_ms":6.953682983294129,"wave_count":1},"fallback_work":448368288,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1143.2077026367188,"n":16,"spread_pct":2.6539115866444316},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.142967915162444,"wave_count":2},"fallback_work":318837056,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":1188.5435791015625,"n":16,"spread_pct":60.592259125897684},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.077852867543697,"wave_count":2},"fallback_work":320806336,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":1195.5195922851562,"n":16,"spread_pct":60.61218049793558},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.149990018457174,"wave_count":2},"fallback_work":331588544,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":1192.9594116210938,"n":16,"spread_pct":2.8266941144912163},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.204765988513827,"wave_count":2},"fallback_work":335871808,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":1189.3521118164062,"n":16,"spread_pct":57.61200471407029},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.1338001396507025,"wave_count":2},"fallback_work":338038016,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":1202.4500122070312,"n":16,"spread_pct":53.050866961126374},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.53536687605083,"wave_count":2},"fallback_work":349607936,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":1177.2117919921875,"n":16,"spread_pct":60.2533939448049},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.95782813616097,"wave_count":2},"fallback_work":361965248,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":1198.6068725585938,"n":16,"spread_pct":2.742044431770587},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.945920944213867,"wave_count":2},"fallback_work":365706880,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":1207.4166259765625,"n":16,"spread_pct":8.81891774642744},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.606653798371553,"wave_count":2},"fallback_work":380132256,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":1202.0061645507812,"n":16,"spread_pct":3.6972287896176175},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.782620050013065,"wave_count":1},"fallback_work":300175168,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":1074.289794921875,"n":16,"spread_pct":69.06676172098618},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":7.411984959617257,"wave_count":1},"fallback_work":383676960,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":1116.7656860351562,"n":16,"spread_pct":59.612837867269484},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.655277986079454,"wave_count":1},"fallback_work":301652608,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":1070.8787841796875,"n":16,"spread_pct":3.295165799025812},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.28394291177392,"wave_count":1},"fallback_work":304902880,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":1073.0328369140625,"n":16,"spread_pct":5.285221127642619},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.994299124926329,"wave_count":1},"fallback_work":397363456,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":1115.2861328125,"n":16,"spread_pct":63.453841459196724},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":5.763168912380934,"wave_count":2},"fallback_work":660213664,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":1330.46728515625,"n":16,"spread_pct":3.3539012264032877},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":6.478278897702694,"wave_count":1},"fallback_work":868514176,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":1305.646728515625,"n":16,"spread_pct":3.5099474955632357},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":7.130590965971351,"wave_count":1},"fallback_work":1088778144,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":1426.1896362304688,"n":16,"spread_pct":4.706879549183922},{"cp_plan":{"max_rank_tokens":16896,"summary_ms":7.8760499600321054,"wave_count":1},"fallback_work":1316033056,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":1601.3550415039062,"n":16,"spread_pct":1.1816774589055952},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":8.360128849744797,"wave_count":1},"fallback_work":1546340352,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":1780.6480102539062,"n":16,"spread_pct":1.166786870165174},{"cp_plan":{"max_rank_tokens":20992,"summary_ms":7.86227616481483,"wave_count":1},"fallback_work":1774087584,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":2000.7631225585938,"n":16,"spread_pct":0.9358123638310247},{"cp_plan":{"max_rank_tokens":23040,"summary_ms":6.692474940791726,"wave_count":1},"fallback_work":2010204256,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":2210.5399169921875,"n":16,"spread_pct":28.276794253471927},{"cp_plan":{"max_rank_tokens":26112,"summary_ms":6.378591991961002,"wave_count":2},"fallback_work":2243563936,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":2473.3848876953125,"n":16,"spread_pct":27.98264077888975},{"cp_plan":{"max_rank_tokens":27648,"summary_ms":7.649122970178723,"wave_count":1},"fallback_work":2480074464,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":2668.906982421875,"n":16,"spread_pct":0.6042439504482456},{"cp_plan":{"max_rank_tokens":30720,"summary_ms":10.223031044006348,"wave_count":1},"fallback_work":2726726784,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":2939.847900390625,"n":16,"spread_pct":0.5109280589143468},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.331911077722907,"wave_count":1},"fallback_work":416810176,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":1115.6688232421875,"n":16,"spread_pct":66.33387891035005},{"cp_plan":{"max_rank_tokens":32768,"summary_ms":8.990037953481078,"wave_count":1},"fallback_work":2979139328,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":3183.5025634765625,"n":16,"spread_pct":0.6904108794833963},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.118414854630828,"wave_count":1},"fallback_work":310565120,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":1084.031005859375,"n":16,"spread_pct":4.161014695088567},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.460606887936592,"wave_count":1},"fallback_work":314257920,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":1077.9132690429688,"n":16,"spread_pct":3.6253609815016317},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.187800852581859,"wave_count":1},"fallback_work":315144096,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":1076.9343872070312,"n":16,"spread_pct":3.3313283640397033},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.065133096650243,"wave_count":1},"fallback_work":315685728,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":1075.5443115234375,"n":16,"spread_pct":3.143144158815843},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.416132044047117,"wave_count":1},"fallback_work":314553312,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":1077.660888671875,"n":16,"spread_pct":6.687378330209631},{"cp_plan":{"max_rank_tokens":35328,"summary_ms":7.470150012522936,"wave_count":1},"fallback_work":3243663264,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":3473.3778076171875,"n":16,"spread_pct":0.564850294217902},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.382977055385709,"wave_count":1},"fallback_work":312532960,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":1083.14794921875,"n":16,"spread_pct":3.1637475222998646},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.375152919441462,"wave_count":1},"fallback_work":304852288,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":1072.865234375,"n":16,"spread_pct":2.6345468016974114},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.776427919045091,"wave_count":1},"fallback_work":298943968,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":1076.1408081054688,"n":16,"spread_pct":63.8605095848105},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.094491109251976,"wave_count":1},"fallback_work":427542752,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1121.4909057617188,"n":16,"spread_pct":62.331326767025644},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.560840876773,"wave_count":1},"fallback_work":401252384,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1117.2249145507812,"n":16,"spread_pct":63.522435574027156},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.506009958684444,"wave_count":1},"fallback_work":384266784,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":1100.8043212890625,"n":16,"spread_pct":64.91142119204608},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.145149003714323,"wave_count":2},"fallback_work":372352160,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":1204.5661010742188,"n":16,"spread_pct":4.2297854370964245},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.508678197860718,"wave_count":2},"fallback_work":361963808,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":1205.8273315429688,"n":16,"spread_pct":2.9868586056835107},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.4630149863660336,"wave_count":2},"fallback_work":347489120,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":1195.4511108398438,"n":16,"spread_pct":3.7842936003489385},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.5032681450247765,"wave_count":2},"fallback_work":338922272,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":1200.5554809570312,"n":16,"spread_pct":3.49138513711889},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.305439932271838,"wave_count":2},"fallback_work":325826080,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":1193.9706420898438,"n":16,"spread_pct":3.1386848440715336}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.807360099628568,"wave_count":2},"fallback_work":606716224,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1369.62744140625,"n":16,"spread_pct":1.9412498480390532},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":9.640182135626674,"wave_count":1},"fallback_work":1161659328,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":1706.4434814453125,"n":16,"spread_pct":1.256246340541312},{"cp_plan":{"max_rank_tokens":20480,"summary_ms":5.87460701353848,"wave_count":1},"fallback_work":1719900896,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":2323.5374755859375,"n":16,"spread_pct":28.94631036544903},{"cp_plan":{"max_rank_tokens":26624,"summary_ms":9.104985976591706,"wave_count":1},"fallback_work":2274204064,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":2990.3282470703125,"n":16,"spread_pct":0.5008660529884573},{"cp_plan":{"max_rank_tokens":31232,"summary_ms":6.5255980007350445,"wave_count":1},"fallback_work":2826784032,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":3589.52783203125,"n":16,"spread_pct":19.17266114142526},{"cp_plan":{"max_rank_tokens":36864,"summary_ms":6.51195808313787,"wave_count":1},"fallback_work":3384582592,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":4305.320068359375,"n":16,"spread_pct":16.21214804471507},{"cp_plan":{"max_rank_tokens":43008,"summary_ms":13.001636136323214,"wave_count":1},"fallback_work":3949076704,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":4968.3798828125,"n":16,"spread_pct":13.245778331365484},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.543013965710998,"wave_count":2},"fallback_work":588153760,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":1351.5859985351562,"n":16,"spread_pct":3.3892835587892822},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.632219137623906,"wave_count":2},"fallback_work":600217120,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1351.4064331054688,"n":16,"spread_pct":1.560440301926121},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.403358886018395,"wave_count":2},"fallback_work":588498944,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1356.95068359375,"n":16,"spread_pct":3.7121428841306425},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.416107062250376,"wave_count":2},"fallback_work":587759424,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":1350.777587890625,"n":16,"spread_pct":2.169513183685974},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.67219702154398,"wave_count":2},"fallback_work":589976384,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1342.3981323242188,"n":16,"spread_pct":51.181678006275476},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.54652395658195,"wave_count":2},"fallback_work":592044128,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1342.7691040039062,"n":16,"spread_pct":49.64243293694266},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.594170885160565,"wave_count":2},"fallback_work":594309280,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1350.1278076171875,"n":16,"spread_pct":55.81237923529194},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.554394960403442,"wave_count":2},"fallback_work":587710752,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":1357.3021850585938,"n":16,"spread_pct":1.7682584332188642},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.615554051473737,"wave_count":2},"fallback_work":593915424,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1343.3386840820312,"n":16,"spread_pct":50.71545110033454},{"cp_plan":{"max_rank_tokens":49280,"summary_ms":9.153911843895912,"wave_count":1},"fallback_work":4513768224,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5690.126953125,"n":16,"spread_pct":11.647659305712162},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.269481964409351,"wave_count":2},"fallback_work":603761824,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1357.1838989257812,"n":16,"spread_pct":2.5619869238766273},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.158287869766355,"wave_count":2},"fallback_work":603564416,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1356.5682373046875,"n":16,"spread_pct":2.1658404426452718},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.840081954374909,"wave_count":2},"fallback_work":599329984,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1356.8744506835938,"n":16,"spread_pct":52.82809420759658},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.821858067065477,"wave_count":2},"fallback_work":592240096,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":1348.3583374023438,"n":16,"spread_pct":53.65708254436106},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.564316106960177,"wave_count":2},"fallback_work":587907200,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":1357.045654296875,"n":16,"spread_pct":2.424071131945023},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.484687095507979,"wave_count":2},"fallback_work":587955872,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":1357.9837036132812,"n":16,"spread_pct":48.589745116735635}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.611465942114592,"wave_count":2},"fallback_work":601777600,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1438.4526977539062,"n":16,"spread_pct":57.379923436441935},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":12.854021042585373,"wave_count":1},"fallback_work":1178875104,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":1827.8814086914062,"n":16,"spread_pct":2.8489306250709356},{"cp_plan":{"max_rank_tokens":20480,"summary_ms":5.856162868440151,"wave_count":1},"fallback_work":1762372768,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":2372.74951171875,"n":16,"spread_pct":33.74433133699264},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.786607973277569,"wave_count":2},"fallback_work":599610912,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1447.524169921875,"n":16,"spread_pct":58.82478147038172},{"cp_plan":{"max_rank_tokens":27136,"summary_ms":10.531788226217031,"wave_count":1},"fallback_work":2346067840,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3117.1363525390625,"n":16,"spread_pct":25.141421297600395},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":9.715633001178503,"wave_count":2},"fallback_work":597641632,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1444.137451171875,"n":16,"spread_pct":4.889473485791743},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.166145903989673,"wave_count":2},"fallback_work":597690464,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1441.759521484375,"n":16,"spread_pct":57.72717966988782}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.801514886319637,"wave_count":2},"fallback_work":587468032,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1438.930908203125,"n":16,"spread_pct":64.38162521986861},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":10.342309949919581,"wave_count":1},"fallback_work":1168504576,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1818.326171875,"n":16,"spread_pct":1.8578548158237045}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.903838831931353,"wave_count":2},"fallback_work":597741856,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1432.0985107421875,"n":16,"spread_pct":58.47263550002042}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13824,"summary_ms":9.517256170511246,"wave_count":1},"fallback_work":1020265824,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":1819.5137329101562,"n":16,"spread_pct":3.7237638444686345},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.583456182852387,"wave_count":2},"fallback_work":796702752,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":1623.657470703125,"n":16,"spread_pct":51.24178614513403},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":6.078273989260197,"wave_count":2},"fallback_work":728909888,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1572.5589599609375,"n":16,"spread_pct":54.573032686525956},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.566652096807957,"wave_count":1},"fallback_work":1082691520,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":1643.230712890625,"n":16,"spread_pct":9.089804243944105},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":8.049055933952332,"wave_count":4},"fallback_work":1355141408,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":1833.4242553710938,"n":16,"spread_pct":4.247776286041567},{"cp_plan":{"max_rank_tokens":17835,"summary_ms":5.754208890721202,"wave_count":1},"fallback_work":1736197568,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":2127.2325439453125,"n":16,"spread_pct":61.10495470378117},{"cp_plan":{"max_rank_tokens":24649,"summary_ms":8.266611024737358,"wave_count":1},"fallback_work":2273023776,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":2815.878662109375,"n":16,"spread_pct":1.2895966382051978},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":7.31314392760396,"wave_count":3},"fallback_work":769723136,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1613.8731079101562,"n":16,"spread_pct":53.99667048601305},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":7.489628158509731,"wave_count":3},"fallback_work":715567456,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1582.1412353515625,"n":16,"spread_pct":6.160968284535158}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11776,"summary_ms":6.533307954668999,"wave_count":2},"fallback_work":827884032,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":1622.0096435546875,"n":16,"spread_pct":50.51087476706496},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":6.72754505649209,"wave_count":1},"fallback_work":934766784,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":1609.1774291992188,"n":16,"spread_pct":2.9669980344598583},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":9.192069992423058,"wave_count":1},"fallback_work":1488873344,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":2048.689208984375,"n":16,"spread_pct":1.3521770005616445},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.005957955494523,"wave_count":2},"fallback_work":604419584,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":1436.1051025390625,"n":16,"spread_pct":59.3668890490118}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10240,"summary_ms":6.341072963550687,"wave_count":2},"fallback_work":644970208,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1478.9338989257812,"n":16,"spread_pct":4.833490591519828},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":8.10830108821392,"wave_count":1},"fallback_work":1232504896,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":1757.337646484375,"n":16,"spread_pct":2.4951740435437073},{"cp_plan":{"max_rank_tokens":21504,"summary_ms":6.451479159295559,"wave_count":1},"fallback_work":1808716224,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":2518.712158203125,"n":16,"spread_pct":30.179947066124697},{"cp_plan":{"max_rank_tokens":25850,"summary_ms":8.266530930995941,"wave_count":1},"fallback_work":2408066592,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":3221.39501953125,"n":16,"spread_pct":23.215813097762354},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.824723048135638,"wave_count":2},"fallback_work":636452672,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1472.9735717773438,"n":16,"spread_pct":53.95892189841886},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.648366874083877,"wave_count":2},"fallback_work":627738528,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1473.3754272460938,"n":16,"spread_pct":3.941236754282339},{"cp_plan":{"max_rank_tokens":33280,"summary_ms":8.481922093778849,"wave_count":1},"fallback_work":3007220512,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3832.46044921875,"n":16,"spread_pct":0.4593968132264493},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.839922931045294,"wave_count":2},"fallback_work":642163584,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1473.1077880859375,"n":16,"spread_pct":52.13760303434813},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.733567988499999,"wave_count":2},"fallback_work":626999648,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1477.022705078125,"n":16,"spread_pct":57.980176482693466},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.725962109863758,"wave_count":2},"fallback_work":627245248,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1474.6064453125,"n":16,"spread_pct":58.347003338417224}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7680,"summary_ms":7.320292992517352,"wave_count":1},"fallback_work":448368288,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1317.2286376953125,"n":16,"spread_pct":3.577429419325112},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.230120150372386,"wave_count":2},"fallback_work":318837056,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":1315.287353515625,"n":16,"spread_pct":66.96142140101648},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.1338039338588715,"wave_count":2},"fallback_work":320806336,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":1333.5148315429688,"n":16,"spread_pct":65.53227213440383},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.316750077530742,"wave_count":2},"fallback_work":331588544,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":1328.50439453125,"n":16,"spread_pct":4.545401920113261},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.364387925714254,"wave_count":2},"fallback_work":335871808,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":1327.5123291015625,"n":16,"spread_pct":63.73703089688156},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.282059010118246,"wave_count":2},"fallback_work":338038016,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":1328.0986328125,"n":16,"spread_pct":58.947567341410455},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.6620958968997,"wave_count":2},"fallback_work":349607936,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":1321.2066650390625,"n":16,"spread_pct":56.69760447921502},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.053827935829759,"wave_count":2},"fallback_work":361965248,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":1322.7205810546875,"n":16,"spread_pct":3.7260813233851318},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.139343837276101,"wave_count":2},"fallback_work":365706880,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":1339.5777587890625,"n":16,"spread_pct":57.765713510362524},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.488112965598702,"wave_count":2},"fallback_work":380132256,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":1342.5106201171875,"n":16,"spread_pct":59.008876919273376},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.825110943987966,"wave_count":1},"fallback_work":300175168,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":1244.2684936523438,"n":16,"spread_pct":71.144777849804},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":7.6935559045523405,"wave_count":1},"fallback_work":383676960,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":1302.6267700195312,"n":16,"spread_pct":60.225767336385594},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.7013139594346285,"wave_count":1},"fallback_work":301652608,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":1246.11083984375,"n":16,"spread_pct":6.34125315679448},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.270839901641011,"wave_count":1},"fallback_work":304902880,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":1245.2008056640625,"n":16,"spread_pct":64.34763939872666},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":7.14639388024807,"wave_count":1},"fallback_work":397363456,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":1295.6032104492188,"n":16,"spread_pct":5.0509528758035165},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":6.060935091227293,"wave_count":2},"fallback_work":660213664,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":1431.8322143554688,"n":16,"spread_pct":56.89162760832782},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":6.883052876219153,"wave_count":1},"fallback_work":868514176,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":1450.4923095703125,"n":16,"spread_pct":56.56253226927516},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":7.279122015461326,"wave_count":1},"fallback_work":1088778144,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":1510.9944458007812,"n":16,"spread_pct":5.625814015460331},{"cp_plan":{"max_rank_tokens":16896,"summary_ms":7.844386855140328,"wave_count":1},"fallback_work":1316033056,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":1707.9170532226562,"n":16,"spread_pct":44.174002527935926},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":8.550720987841487,"wave_count":1},"fallback_work":1546340352,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":1841.785400390625,"n":16,"spread_pct":45.20592850989395},{"cp_plan":{"max_rank_tokens":20992,"summary_ms":8.129335008561611,"wave_count":1},"fallback_work":1774087584,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":2059.8297119140625,"n":16,"spread_pct":37.66366552339751},{"cp_plan":{"max_rank_tokens":23040,"summary_ms":6.9735979195684195,"wave_count":1},"fallback_work":2010204256,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":2239.6536865234375,"n":16,"spread_pct":1.3103113094430592},{"cp_plan":{"max_rank_tokens":26112,"summary_ms":8.68670898489654,"wave_count":2},"fallback_work":2243563936,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":2524.2967529296875,"n":16,"spread_pct":5.374988653969848},{"cp_plan":{"max_rank_tokens":27648,"summary_ms":8.092141011729836,"wave_count":1},"fallback_work":2480074464,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":2672.6275634765625,"n":16,"spread_pct":27.74805067925632},{"cp_plan":{"max_rank_tokens":30720,"summary_ms":10.639207903295755,"wave_count":1},"fallback_work":2726726784,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":2956.695068359375,"n":16,"spread_pct":0.6126522145062232},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.607310147956014,"wave_count":1},"fallback_work":416810176,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":1295.3868408203125,"n":16,"spread_pct":60.0071769101762},{"cp_plan":{"max_rank_tokens":32768,"summary_ms":9.236284997314215,"wave_count":1},"fallback_work":2979139328,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":3153.886962890625,"n":16,"spread_pct":0.6865752910149856},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.233773125335574,"wave_count":1},"fallback_work":310565120,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":1253.251220703125,"n":16,"spread_pct":3.7758431828776597},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.536859156563878,"wave_count":1},"fallback_work":314257920,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":1262.7073974609375,"n":16,"spread_pct":68.45077565479603},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.563033977523446,"wave_count":1},"fallback_work":315144096,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":1259.0562133789062,"n":16,"spread_pct":6.525670924086757},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.323884077370167,"wave_count":1},"fallback_work":315685728,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":1257.59619140625,"n":16,"spread_pct":70.07599715323714},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.523510044440627,"wave_count":1},"fallback_work":314553312,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":1258.1249389648438,"n":16,"spread_pct":9.206651356352234},{"cp_plan":{"max_rank_tokens":35328,"summary_ms":7.389484904706478,"wave_count":1},"fallback_work":3243663264,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":3399.0784912109375,"n":16,"spread_pct":23.349733855873662},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.506352849304676,"wave_count":1},"fallback_work":312532960,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":1258.9818115234375,"n":16,"spread_pct":4.314351012388619},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.446049151942134,"wave_count":1},"fallback_work":304852288,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":1250.4644165039062,"n":16,"spread_pct":65.73673872827229},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.893642082810402,"wave_count":1},"fallback_work":298943968,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":1244.704345703125,"n":16,"spread_pct":9.510319990412508},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.195612018927932,"wave_count":1},"fallback_work":427542752,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1301.0360107421875,"n":16,"spread_pct":2.6307629127065675},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.755919894203544,"wave_count":1},"fallback_work":401252384,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1302.892578125,"n":16,"spread_pct":2.6721455115645627},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.523583084344864,"wave_count":1},"fallback_work":384266784,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":1282.4736328125,"n":16,"spread_pct":63.6732031070936},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.309080868959427,"wave_count":2},"fallback_work":372352160,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":1339.6373901367188,"n":16,"spread_pct":59.5879768107506},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":8.55671800673008,"wave_count":2},"fallback_work":361963808,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":1337.3389892578125,"n":16,"spread_pct":3.4340345257319145},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.910970900207758,"wave_count":2},"fallback_work":347489120,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":1327.503662109375,"n":16,"spread_pct":62.05100547958558},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.654229946434498,"wave_count":2},"fallback_work":338922272,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":1334.2561645507812,"n":16,"spread_pct":4.510976968318086},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.496144061908126,"wave_count":2},"fallback_work":325826080,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":1324.3065185546875,"n":16,"spread_pct":67.00962574475383}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.884522968903184,"wave_count":2},"fallback_work":606716224,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1492.5855102539062,"n":16,"spread_pct":51.27745499492712},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":9.567287983372808,"wave_count":1},"fallback_work":1161659328,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":1817.24072265625,"n":16,"spread_pct":46.44117124686535},{"cp_plan":{"max_rank_tokens":20480,"summary_ms":5.859010154381394,"wave_count":1},"fallback_work":1719900896,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":2374.1676025390625,"n":16,"spread_pct":5.4712432364118575},{"cp_plan":{"max_rank_tokens":26624,"summary_ms":9.286504006013274,"wave_count":1},"fallback_work":2274204064,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":3031.1204833984375,"n":16,"spread_pct":25.434140851459546},{"cp_plan":{"max_rank_tokens":31232,"summary_ms":6.810406921431422,"wave_count":1},"fallback_work":2826784032,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":3535.9779052734375,"n":16,"spread_pct":44.002093297997625},{"cp_plan":{"max_rank_tokens":36864,"summary_ms":6.942945998162031,"wave_count":1},"fallback_work":3384582592,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":4204.26953125,"n":16,"spread_pct":17.25892716946036},{"cp_plan":{"max_rank_tokens":43008,"summary_ms":8.132823975756764,"wave_count":1},"fallback_work":3949076704,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":4893.241455078125,"n":16,"spread_pct":16.598857330570503},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.4770239405334,"wave_count":2},"fallback_work":588153760,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":1472.247802734375,"n":16,"spread_pct":2.262215631092642},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.643210140988231,"wave_count":2},"fallback_work":600217120,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1480.4417724609375,"n":16,"spread_pct":58.52642928272357},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.329204887151718,"wave_count":2},"fallback_work":588498944,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1480.5675659179688,"n":16,"spread_pct":53.086042781033036},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.4660049974918365,"wave_count":2},"fallback_work":587759424,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":1482.38037109375,"n":16,"spread_pct":2.9831780534039987},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.7064950708299875,"wave_count":2},"fallback_work":589976384,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1468.8872680664062,"n":16,"spread_pct":53.66231355787934},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.673198029398918,"wave_count":2},"fallback_work":592044128,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1466.7725219726562,"n":16,"spread_pct":54.99019395336998},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.587287014350295,"wave_count":2},"fallback_work":594309280,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1464.290283203125,"n":16,"spread_pct":3.2788305181307673},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.485790014266968,"wave_count":2},"fallback_work":587710752,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":1481.3908081054688,"n":16,"spread_pct":4.422863777906793},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.671768914908171,"wave_count":2},"fallback_work":593915424,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1472.4185180664062,"n":16,"spread_pct":2.640064781677911},{"cp_plan":{"max_rank_tokens":49280,"summary_ms":9.241929976269603,"wave_count":1},"fallback_work":4513768224,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5584.197265625,"n":16,"spread_pct":13.94894353450673},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.171663990244269,"wave_count":2},"fallback_work":603761824,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1483.152587890625,"n":16,"spread_pct":3.3608927615225235},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.063921842724085,"wave_count":2},"fallback_work":603564416,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1476.7369384765625,"n":16,"spread_pct":3.5213184553511327},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.809505004435778,"wave_count":2},"fallback_work":599329984,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1478.3366088867188,"n":16,"spread_pct":6.756622241291021},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.727954907342792,"wave_count":2},"fallback_work":592240096,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":1468.4296875,"n":16,"spread_pct":2.7271655653626588},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.567102856934071,"wave_count":2},"fallback_work":587907200,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":1484.7671508789062,"n":16,"spread_pct":3.9682608521391862},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.4820770639926195,"wave_count":2},"fallback_work":587955872,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":1480.59326171875,"n":16,"spread_pct":3.397586944369784}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.6388691514730453,"wave_count":2},"fallback_work":601877920,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1200.5313110351562,"n":8,"spread_pct":344.9220201525153},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":6.996108219027519,"wave_count":3},"fallback_work":1178694208,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":1435.9537353515625,"n":8,"spread_pct":49.08122553361364},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":10.219024494290352,"wave_count":1},"fallback_work":1761907424,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":1460.3027954101562,"n":8,"spread_pct":2.6220114596076396},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.5911835730075836,"wave_count":2},"fallback_work":599843616,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1216.1848754882812,"n":8,"spread_pct":6.201845199730021},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":10.522767901420593,"wave_count":1},"fallback_work":2345186624,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1826.6144409179688,"n":8,"spread_pct":2.2987009398223712},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.966415464878082,"wave_count":2},"fallback_work":597875296,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1236.2926025390625,"n":8,"spread_pct":298.0828734316912},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.964249208569527,"wave_count":2},"fallback_work":598055296,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1234.5927734375,"n":8,"spread_pct":60.18675107160683}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.4945718944072723,"wave_count":2},"fallback_work":587444096,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1207.7838134765625,"n":8,"spread_pct":3.8172801848362146},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":8.71739536523819,"wave_count":3},"fallback_work":1168066176,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1352.2539672851562,"n":8,"spread_pct":53.43208399207252}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.3876933157444,"wave_count":2},"fallback_work":597516192,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1220.1919555664062,"n":8,"spread_pct":3.863807969587265}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":8192,"summary_ms":6.507227197289467,"wave_count":1},"fallback_work":1020227840,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":1270.3474731445312,"n":8,"spread_pct":57.560819190623256},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":3.868779167532921,"wave_count":2},"fallback_work":796904992,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":1290.4193725585938,"n":8,"spread_pct":6.427438996385954},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.7754345685243607,"wave_count":2},"fallback_work":729276256,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1283.5657348632812,"n":8,"spread_pct":4.287848453783108},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.276204854249954,"wave_count":2},"fallback_work":1082754208,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":1281.3963012695312,"n":8,"spread_pct":298.2630365616121},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":4.507711157202721,"wave_count":1},"fallback_work":1355071392,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":1407.276123046875,"n":8,"spread_pct":258.43829580602903},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":5.086807534098625,"wave_count":2},"fallback_work":1735810464,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":1417.5327758789062,"n":8,"spread_pct":49.294512340807955},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":7.942821830511093,"wave_count":1},"fallback_work":2272243872,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":1686.5443115234375,"n":8,"spread_pct":13.162047574541456},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":5.8549195528030396,"wave_count":3},"fallback_work":770069632,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1378.9837036132812,"n":8,"spread_pct":49.92395744059879},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":10.34148596227169,"wave_count":2},"fallback_work":716071680,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1287.3614501953125,"n":8,"spread_pct":66.0472223526026}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7680,"summary_ms":6.764085963368416,"wave_count":2},"fallback_work":827808512,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":1284.5186767578125,"n":8,"spread_pct":301.5531831568796},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.651885807514191,"wave_count":4},"fallback_work":934639136,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":1376.3687133789062,"n":8,"spread_pct":53.38398721975306},{"cp_plan":{"max_rank_tokens":12288,"summary_ms":9.01821255683899,"wave_count":1},"fallback_work":1488344384,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":1382.8267211914062,"n":8,"spread_pct":324.5110341158656},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.245348274707794,"wave_count":2},"fallback_work":604584192,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":1219.4669799804688,"n":8,"spread_pct":6.080877800905086}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.101839244365692,"wave_count":4},"fallback_work":645115200,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1461.1905517578125,"n":8,"spread_pct":52.63456622194368},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":4.268443211913109,"wave_count":2},"fallback_work":1232363360,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":1343.7030639648438,"n":8,"spread_pct":49.54886005965597},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":8.261928334832191,"wave_count":2},"fallback_work":1808293792,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":1582.6884765625,"n":8,"spread_pct":47.17409064012296},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.285088136792183,"wave_count":1},"fallback_work":2407352096,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":1832.0997314453125,"n":8,"spread_pct":1.8771686372864127},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.6718249320983887,"wave_count":2},"fallback_work":636732896,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1226.9920043945312,"n":8,"spread_pct":2.252525327392075},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.1632745414972305,"wave_count":2},"fallback_work":628023136,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1233.8153686523438,"n":8,"spread_pct":56.48057112127229},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":14.572562649846077,"wave_count":3},"fallback_work":3006082720,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2136.2806396484375,"n":8,"spread_pct":5.18036976388726},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.69291752576828,"wave_count":2},"fallback_work":642441024,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1235.1234741210938,"n":8,"spread_pct":58.80911905523267},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.231764003634453,"wave_count":2},"fallback_work":627415808,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1244.0537109375,"n":8,"spread_pct":59.97497077896337},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.32843342423439,"wave_count":2},"fallback_work":627792416,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1243.9375610351562,"n":8,"spread_pct":4.77181494448196}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":4096,"summary_ms":3.7964023649692535,"wave_count":1},"fallback_work":449134048,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1114.4505615234375,"n":8,"spread_pct":64.16254243487022},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6335326731204987,"wave_count":2},"fallback_work":320190688,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":1215.1239013671875,"n":8,"spread_pct":57.74561633947959},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.620695322751999,"wave_count":2},"fallback_work":322159008,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":1181.795166015625,"n":8,"spread_pct":74.57658946670817},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.618786111474037,"wave_count":2},"fallback_work":332804768,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":1207.6091918945312,"n":8,"spread_pct":58.95299221868332},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.5965777933597565,"wave_count":2},"fallback_work":337085920,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":1189.9169311523438,"n":8,"spread_pct":6.478056919410376},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.602122887969017,"wave_count":2},"fallback_work":339251072,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":1193.23046875,"n":8,"spread_pct":5.357411766246436},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.5675447434186935,"wave_count":1},"fallback_work":350684160,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":1080.09228515625,"n":8,"spread_pct":6.804216852491536},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.60182486474514,"wave_count":2},"fallback_work":363035424,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":1205.1622314453125,"n":8,"spread_pct":11.60182398128818},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.5953838378190994,"wave_count":2},"fallback_work":366775232,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":1222.4283447265625,"n":8,"spread_pct":54.84844269593783},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.564363345503807,"wave_count":1},"fallback_work":381062496,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":1082.8841552734375,"n":8,"spread_pct":4.087256581482985},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.899887204170227,"wave_count":2},"fallback_work":302325216,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":1180.9212036132812,"n":8,"spread_pct":62.053182896012146},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.607584163546562,"wave_count":1},"fallback_work":384605472,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":1115.8597412109375,"n":8,"spread_pct":388.0952701325647},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.8193687796592712,"wave_count":2},"fallback_work":303670720,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":1188.6386108398438,"n":8,"spread_pct":6.697073709393055},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.714868798851967,"wave_count":2},"fallback_work":306657088,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":1177.9859619140625,"n":8,"spread_pct":7.5015069852926315},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.5621747374534607,"wave_count":2},"fallback_work":398285184,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":1244.0309448242188,"n":8,"spread_pct":303.53504967053857},{"cp_plan":{"max_rank_tokens":4608,"summary_ms":3.820352256298065,"wave_count":3},"fallback_work":660876128,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":1246.6842651367188,"n":8,"spread_pct":3.1554893534730186},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.140174016356468,"wave_count":1},"fallback_work":869075008,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":1180.498779296875,"n":8,"spread_pct":6.125929953386774},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.3857768177986145,"wave_count":1},"fallback_work":1089231712,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":1212.3622436523438,"n":8,"spread_pct":9.807180996673118},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":4.8485323786735535,"wave_count":3},"fallback_work":1316375840,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":1306.4058837890625,"n":8,"spread_pct":6.229108611347124},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":5.04249706864357,"wave_count":1},"fallback_work":1546570752,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":1311.9491577148438,"n":8,"spread_pct":3.3881034986083955},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.370140075683594,"wave_count":1},"fallback_work":1774207072,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":1372.737060546875,"n":8,"spread_pct":9.403393119820244},{"cp_plan":{"max_rank_tokens":12288,"summary_ms":5.785055458545685,"wave_count":3},"fallback_work":2010208640,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":1475.9843139648438,"n":8,"spread_pct":9.456794982592182},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":9.71284694969654,"wave_count":3},"fallback_work":2243454560,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":1499.8413696289062,"n":8,"spread_pct":2.748524715809586},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.595641374588013,"wave_count":3},"fallback_work":2479849792,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":1567.7960205078125,"n":8,"spread_pct":145.670164057217},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":6.899183616042137,"wave_count":1},"fallback_work":2726381760,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":1665.608154296875,"n":8,"spread_pct":233.4884293945867},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6145560443401337,"wave_count":1},"fallback_work":417722400,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":1091.9012451171875,"n":8,"spread_pct":148.79194799692112},{"cp_plan":{"max_rank_tokens":16896,"summary_ms":7.317868992686272,"wave_count":1},"fallback_work":2978671232,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":1739.6707153320312,"n":8,"spread_pct":184.5837693290458},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.7486832588911057,"wave_count":1},"fallback_work":312185216,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":1075.0322875976562,"n":8,"spread_pct":4.788839179691698},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.720685839653015,"wave_count":2},"fallback_work":315745024,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":1182.7523193359375,"n":8,"spread_pct":7.359759779719638},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6819949746131897,"wave_count":2},"fallback_work":316630880,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":1185.0279541015625,"n":8,"spread_pct":7.119436235938939},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6663487553596497,"wave_count":2},"fallback_work":317172224,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":1170.429443359375,"n":8,"spread_pct":3.833718635007892},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6301910877227783,"wave_count":2},"fallback_work":316040384,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":1171.9112548828125,"n":8,"spread_pct":20.414003858422298},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":6.182560697197914,"wave_count":1},"fallback_work":3242935136,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":1846.413818359375,"n":8,"spread_pct":2.2905333673766624},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.916781395673752,"wave_count":1},"fallback_work":314545856,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":1078.8704223632812,"n":8,"spread_pct":3.0792681667730526},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.985527902841568,"wave_count":2},"fallback_work":307000032,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":1194.0137939453125,"n":8,"spread_pct":60.338167663988685},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.968711942434311,"wave_count":2},"fallback_work":301225920,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":1186.886962890625,"n":8,"spread_pct":77.69203015783847},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":4.161331802606583,"wave_count":1},"fallback_work":428449856,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1109.6231079101562,"n":8,"spread_pct":78.85379947761463},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.7119146436452866,"wave_count":2},"fallback_work":402303520,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1209.7040405273438,"n":8,"spread_pct":58.93694842695364},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6935079842805862,"wave_count":1},"fallback_work":385457440,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":1090.1668701171875,"n":8,"spread_pct":3.392956859676572},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":4.319103434681892,"wave_count":2},"fallback_work":373679840,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":1190.65283203125,"n":8,"spread_pct":82.58488872266355},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.805069252848625,"wave_count":2},"fallback_work":363427744,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":1204.3556518554688,"n":8,"spread_pct":4.724692371535752},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.803761675953865,"wave_count":2},"fallback_work":349091328,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":1199.887451171875,"n":8,"spread_pct":113.56739391450898},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.9820633828639984,"wave_count":2},"fallback_work":340659872,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":1200.2484741210938,"n":8,"spread_pct":8.852364198066695},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.00182418525219,"wave_count":2},"fallback_work":327701280,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":1178.9343872070312,"n":8,"spread_pct":7.74093105071579}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.754507750272751,"wave_count":2},"fallback_work":607076576,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1354.9669189453125,"n":8,"spread_pct":9.76216068466991},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":6.096480414271355,"wave_count":3},"fallback_work":1161749152,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":1469.1829833984375,"n":8,"spread_pct":3.532103692503548},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.32536444067955,"wave_count":2},"fallback_work":1719718720,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":1631.5159912109375,"n":8,"spread_pct":5.513236848726848},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":5.968812853097916,"wave_count":1},"fallback_work":2273751648,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":1803.39697265625,"n":8,"spread_pct":51.15606208660576},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":11.708203703165054,"wave_count":3},"fallback_work":2826062240,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":2057.4413452148438,"n":8,"spread_pct":3.2027654000687464},{"cp_plan":{"max_rank_tokens":19456,"summary_ms":7.136747241020203,"wave_count":1},"fallback_work":3383588768,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":2282.4874267578125,"n":8,"spread_pct":1.4597390103623693},{"cp_plan":{"max_rank_tokens":21504,"summary_ms":11.523066088557243,"wave_count":1},"fallback_work":3947807808,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":2594.635498046875,"n":8,"spread_pct":36.76634492347163},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.837594762444496,"wave_count":2},"fallback_work":589048160,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":1355.599365234375,"n":8,"spread_pct":10.633548460297863},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.6733001470565796,"wave_count":2},"fallback_work":600711968,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1327.0864868164062,"n":8,"spread_pct":60.856335963070386},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.741348162293434,"wave_count":2},"fallback_work":589261824,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1358.6040649414062,"n":8,"spread_pct":68.6192137791507},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.310624673962593,"wave_count":2},"fallback_work":588785120,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":1340.6917724609375,"n":8,"spread_pct":11.4613318265125},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.723347559571266,"wave_count":2},"fallback_work":590607328,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1365.0296020507812,"n":8,"spread_pct":138.08886302465342},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.586305469274521,"wave_count":2},"fallback_work":592674176,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1350.4097900390625,"n":8,"spread_pct":65.30954823554808},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.578225314617157,"wave_count":2},"fallback_work":594807008,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1347.9844970703125,"n":8,"spread_pct":4.807995588755601},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.8065649569034576,"wave_count":2},"fallback_work":588605344,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":1349.2255859375,"n":8,"spread_pct":11.44351288029085},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.592845216393471,"wave_count":2},"fallback_work":594413344,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1363.1842041015625,"n":8,"spread_pct":51.70979667696617},{"cp_plan":{"max_rank_tokens":24576,"summary_ms":9.03359055519104,"wave_count":1},"fallback_work":4512092832,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2984.0450439453125,"n":8,"spread_pct":2.046708616385393},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.732386976480484,"wave_count":2},"fallback_work":604254944,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1341.1376953125,"n":8,"spread_pct":27.035060892359787},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7183519452810287,"wave_count":2},"fallback_work":604188736,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1348.8051147460938,"n":8,"spread_pct":58.62492314635942},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.680795595049858,"wave_count":2},"fallback_work":600087584,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1360.621826171875,"n":8,"spread_pct":74.42417667559586},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.735488444566727,"wave_count":2},"fallback_work":593132480,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":1354.2606811523438,"n":8,"spread_pct":8.19522731151733},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.8364920765161514,"wave_count":2},"fallback_work":588932800,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":1352.2444458007812,"n":8,"spread_pct":9.576174376972817},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.8909558206796646,"wave_count":2},"fallback_work":589112800,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":1348.8886108398438,"n":8,"spread_pct":60.12139895080545}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.19561006128788,"wave_count":2},"fallback_work":601877920,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1181.8789672851562,"n":16,"spread_pct":4.357788369649613},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":7.7086330857127905,"wave_count":3},"fallback_work":1178694208,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":1389.4163208007812,"n":16,"spread_pct":48.257059100305625},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":10.108692105859518,"wave_count":1},"fallback_work":1761907424,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":1493.871337890625,"n":16,"spread_pct":4.229692050707832},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.9757308550179005,"wave_count":2},"fallback_work":599843616,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1184.4442138671875,"n":16,"spread_pct":194.72329443148678},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":7.82920396886766,"wave_count":1},"fallback_work":2345186624,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1890.5856323242188,"n":16,"spread_pct":78.34573001665356},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.072302184998989,"wave_count":2},"fallback_work":597875296,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1199.9635620117188,"n":16,"spread_pct":96.10636268608214},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.354863919317722,"wave_count":2},"fallback_work":598055296,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1197.4552612304688,"n":16,"spread_pct":76.02409222805096}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.8820451591163874,"wave_count":2},"fallback_work":587444096,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1149.351806640625,"n":16,"spread_pct":127.36104015829235},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":10.899567976593971,"wave_count":3},"fallback_work":1168066176,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1334.5951538085938,"n":16,"spread_pct":55.3948601589673}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.745427053421736,"wave_count":2},"fallback_work":597516192,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1131.5926513671875,"n":16,"spread_pct":63.12838814469072}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":8192,"summary_ms":7.540390826761723,"wave_count":1},"fallback_work":1020227840,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":1224.5304565429688,"n":16,"spread_pct":124.79807435756301},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":4.142635967582464,"wave_count":2},"fallback_work":796904992,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":1264.6161499023438,"n":16,"spread_pct":57.53925860365836},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.919634036719799,"wave_count":2},"fallback_work":729276256,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1243.5650024414062,"n":16,"spread_pct":6.237157703802225},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.434763919562101,"wave_count":2},"fallback_work":1082754208,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":1253.1539916992188,"n":16,"spread_pct":63.35500634994465},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":4.912347998470068,"wave_count":1},"fallback_work":1355071392,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":1352.3884887695312,"n":16,"spread_pct":49.77690431511486},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":4.72704297862947,"wave_count":2},"fallback_work":1735810464,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":1429.4672241210938,"n":16,"spread_pct":53.90360049924017},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":9.303495986387134,"wave_count":1},"fallback_work":2272243872,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":1711.3285522460938,"n":16,"spread_pct":42.96023664967842},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":6.301112938672304,"wave_count":3},"fallback_work":770069632,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1328.158203125,"n":16,"spread_pct":50.30519457601795},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":10.942380176857114,"wave_count":2},"fallback_work":716071680,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1229.3300170898438,"n":16,"spread_pct":66.82693403314765}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7680,"summary_ms":7.760311011224985,"wave_count":2},"fallback_work":827808512,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":1311.9834594726562,"n":16,"spread_pct":4.678171101284885},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.424559047445655,"wave_count":4},"fallback_work":934639136,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":1401.7725830078125,"n":16,"spread_pct":58.71569731439189},{"cp_plan":{"max_rank_tokens":12288,"summary_ms":11.63804205134511,"wave_count":1},"fallback_work":1488344384,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":1392.049560546875,"n":16,"spread_pct":58.078592242672755},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.952166996896267,"wave_count":2},"fallback_work":604584192,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":1202.18994140625,"n":16,"spread_pct":56.95109023449637}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.927513983100653,"wave_count":4},"fallback_work":645115200,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1390.7476806640625,"n":16,"spread_pct":57.18952989136755},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":5.422478076070547,"wave_count":2},"fallback_work":1232363360,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":1328.100830078125,"n":16,"spread_pct":103.69764082103701},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":8.870910853147507,"wave_count":2},"fallback_work":1808293792,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":1597.1447143554688,"n":16,"spread_pct":62.97708629891011},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.775395013391972,"wave_count":1},"fallback_work":2407352096,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":1959.1099243164062,"n":16,"spread_pct":1.2725754708454784},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.8577590603381395,"wave_count":2},"fallback_work":636732896,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1177.1008911132812,"n":16,"spread_pct":3.5198036920874887},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.463822115212679,"wave_count":2},"fallback_work":628023136,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1190.4948120117188,"n":16,"spread_pct":3.673322424373113},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":20.62675985507667,"wave_count":3},"fallback_work":3006082720,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2264.921875,"n":16,"spread_pct":30.771758907936945},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.889573970809579,"wave_count":2},"fallback_work":642441024,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1185.7984619140625,"n":16,"spread_pct":121.45806990505825},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.518296103924513,"wave_count":2},"fallback_work":627415808,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1191.7465209960938,"n":16,"spread_pct":60.56748648887421},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":9.28241084329784,"wave_count":2},"fallback_work":627792416,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1186.79150390625,"n":16,"spread_pct":4.9626648135873825}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":4096,"summary_ms":4.1560588870197535,"wave_count":1},"fallback_work":449134048,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1058.2145385742188,"n":16,"spread_pct":5.905274992708123},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":8.45855101943016,"wave_count":2},"fallback_work":320190688,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":1141.0987548828125,"n":16,"spread_pct":7.335277374626399},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.817383199930191,"wave_count":2},"fallback_work":322159008,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":1137.1608276367188,"n":16,"spread_pct":55.943288378558535},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6903650034219027,"wave_count":2},"fallback_work":332804768,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":1144.0322265625,"n":16,"spread_pct":65.48229432798772},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6624018102884293,"wave_count":2},"fallback_work":337085920,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":1159.4365844726562,"n":16,"spread_pct":73.4300301707567},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.76377091743052,"wave_count":2},"fallback_work":339251072,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":1156.767822265625,"n":16,"spread_pct":60.48119075051059},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6603661719709635,"wave_count":1},"fallback_work":350684160,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":1033.0322265625,"n":16,"spread_pct":6.81046014227306},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.667111974209547,"wave_count":2},"fallback_work":363035424,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":1165.6611938476562,"n":16,"spread_pct":60.48577475366387},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6428410094231367,"wave_count":2},"fallback_work":366775232,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":1131.6650390625,"n":16,"spread_pct":63.199145686363345},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6216899752616882,"wave_count":1},"fallback_work":381062496,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":1034.2352905273438,"n":16,"spread_pct":62.43359436942454},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.003201145678759,"wave_count":2},"fallback_work":302325216,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":1120.4024047851562,"n":16,"spread_pct":69.07269089556696},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.7114459555596113,"wave_count":1},"fallback_work":384605472,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":1022.4822692871094,"n":16,"spread_pct":65.86605060979656},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.889031009748578,"wave_count":2},"fallback_work":303670720,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":1134.6219482421875,"n":16,"spread_pct":56.927915679513205},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.84885398671031,"wave_count":2},"fallback_work":306657088,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":1144.89208984375,"n":16,"spread_pct":68.8163160164095},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.678248031064868,"wave_count":2},"fallback_work":398285184,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":1146.7218627929688,"n":16,"spread_pct":9.535780253355739},{"cp_plan":{"max_rank_tokens":4608,"summary_ms":4.110265988856554,"wave_count":3},"fallback_work":660876128,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":1191.1808471679688,"n":16,"spread_pct":90.50379254640747},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.4553871266543865,"wave_count":1},"fallback_work":869075008,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":1128.5681762695312,"n":16,"spread_pct":56.70414599181708},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.653580952435732,"wave_count":1},"fallback_work":1089231712,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":1188.8926391601562,"n":16,"spread_pct":62.56973408600738},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":4.946643952280283,"wave_count":3},"fallback_work":1316375840,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":1290.0323486328125,"n":16,"spread_pct":54.88594045483772},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":5.611483007669449,"wave_count":1},"fallback_work":1546570752,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":1307.0161743164062,"n":16,"spread_pct":55.98139697610267},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":4.878271138295531,"wave_count":1},"fallback_work":1774207072,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":1401.839599609375,"n":16,"spread_pct":48.64750297591317},{"cp_plan":{"max_rank_tokens":12288,"summary_ms":5.263072904199362,"wave_count":3},"fallback_work":2010208640,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":1504.2672119140625,"n":16,"spread_pct":43.232902622317035},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":9.741223882883787,"wave_count":3},"fallback_work":2243454560,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":1576.4904174804688,"n":16,"spread_pct":39.80469560461739},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.13308185711503,"wave_count":3},"fallback_work":2479849792,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":1663.9124755859375,"n":16,"spread_pct":38.44116374635008},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":6.833428982645273,"wave_count":1},"fallback_work":2726381760,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":1766.1021728515625,"n":16,"spread_pct":45.0324231373034},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6696521565318108,"wave_count":1},"fallback_work":417722400,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":1032.0668334960938,"n":16,"spread_pct":6.822774077563919},{"cp_plan":{"max_rank_tokens":16896,"summary_ms":7.355111883953214,"wave_count":1},"fallback_work":2978671232,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":1871.6890258789062,"n":16,"spread_pct":2.19735024931558},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.7781349383294582,"wave_count":1},"fallback_work":312185216,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":1021.5628967285156,"n":16,"spread_pct":67.3811373187643},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.776432015001774,"wave_count":2},"fallback_work":315745024,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":1137.4083862304688,"n":16,"spread_pct":63.37791521348812},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.731217933818698,"wave_count":2},"fallback_work":316630880,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":1131.3377075195312,"n":16,"spread_pct":7.489653690847627},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.7041439209133387,"wave_count":2},"fallback_work":317172224,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":1127.2677001953125,"n":16,"spread_pct":6.195811676051865},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.014601930975914,"wave_count":2},"fallback_work":316040384,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":1131.996337890625,"n":16,"spread_pct":6.512817316441566},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":7.253503194078803,"wave_count":1},"fallback_work":3242935136,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":2066.6717529296875,"n":16,"spread_pct":0.7516283795058232},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.964601084589958,"wave_count":1},"fallback_work":314545856,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":1019.9254150390625,"n":16,"spread_pct":66.2171258262341},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.031151998788118,"wave_count":2},"fallback_work":307000032,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":1143.2852783203125,"n":16,"spread_pct":61.50926034773564},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.082807106897235,"wave_count":2},"fallback_work":301225920,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":1122.1050415039062,"n":16,"spread_pct":9.381715720395729},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":3.8219860289245844,"wave_count":1},"fallback_work":428449856,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1054.17822265625,"n":16,"spread_pct":65.85377221500937},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.842010861262679,"wave_count":2},"fallback_work":402303520,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1163.1383056640625,"n":16,"spread_pct":59.61656428357707},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.8542989641427994,"wave_count":1},"fallback_work":385457440,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":1046.8303833007812,"n":16,"spread_pct":65.0809977967228},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.855125978589058,"wave_count":2},"fallback_work":373679840,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":1156.9854736328125,"n":16,"spread_pct":60.423347691106535},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.8913257885724306,"wave_count":2},"fallback_work":363427744,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":1146.9485473632812,"n":16,"spread_pct":59.31260262228798},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":4.006752045825124,"wave_count":2},"fallback_work":349091328,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":1160.1336059570312,"n":16,"spread_pct":5.816022608606261},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.934659995138645,"wave_count":2},"fallback_work":340659872,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":1148.7382202148438,"n":16,"spread_pct":8.88140100644352},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.981512039899826,"wave_count":2},"fallback_work":327701280,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":1137.3641967773438,"n":16,"spread_pct":6.459540933989641}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.389976151287556,"wave_count":2},"fallback_work":607076576,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1190.1771240234375,"n":16,"spread_pct":59.930424123001494},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":6.690181791782379,"wave_count":3},"fallback_work":1161749152,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":1357.4462890625,"n":16,"spread_pct":54.35344688045179},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":10.266430908814073,"wave_count":2},"fallback_work":1719718720,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":1533.2744750976562,"n":16,"spread_pct":2.0156068374047402},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":5.935607943683863,"wave_count":1},"fallback_work":2273751648,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":1824.2704467773438,"n":16,"spread_pct":36.53553447545166},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":14.03611688874662,"wave_count":3},"fallback_work":2826062240,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":2154.5147705078125,"n":16,"spread_pct":1.9545733843216788},{"cp_plan":{"max_rank_tokens":19456,"summary_ms":7.9707009717822075,"wave_count":1},"fallback_work":3383588768,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":2505.7181396484375,"n":16,"spread_pct":0.749545363600238},{"cp_plan":{"max_rank_tokens":21504,"summary_ms":9.850959992036223,"wave_count":1},"fallback_work":3947807808,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":2796.960693359375,"n":16,"spread_pct":44.47887464341831},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.887317841872573,"wave_count":2},"fallback_work":589048160,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":1175.1554565429688,"n":16,"spread_pct":5.902373184814364},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.806163091212511,"wave_count":2},"fallback_work":600711968,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1172.828125,"n":16,"spread_pct":4.489192540067412},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.8582468405365944,"wave_count":2},"fallback_work":589261824,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1177.7029418945312,"n":16,"spread_pct":59.53729590108774},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.9789059665054083,"wave_count":2},"fallback_work":588785120,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":1173.17724609375,"n":16,"spread_pct":58.47961869039696},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.801102051511407,"wave_count":2},"fallback_work":590607328,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1183.489013671875,"n":16,"spread_pct":57.96324592254554},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.860304994508624,"wave_count":2},"fallback_work":592674176,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1186.8684692382812,"n":16,"spread_pct":59.27797847673205},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.781122086569667,"wave_count":2},"fallback_work":594807008,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1185.6946411132812,"n":16,"spread_pct":56.160380751554804},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.9036660455167294,"wave_count":2},"fallback_work":588605344,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":1176.2816772460938,"n":16,"spread_pct":60.502716892867284},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.754270892590284,"wave_count":2},"fallback_work":594413344,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1186.0624389648438,"n":16,"spread_pct":58.11616004922077},{"cp_plan":{"max_rank_tokens":24576,"summary_ms":10.48646098934114,"wave_count":1},"fallback_work":4512092832,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3247.500732421875,"n":16,"spread_pct":0.6324800704989255},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.101301077753305,"wave_count":2},"fallback_work":604254944,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1177.0831909179688,"n":16,"spread_pct":62.0989082323762},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.123274935409427,"wave_count":2},"fallback_work":604188736,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1178.718017578125,"n":16,"spread_pct":65.70277436196012},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.026644095778465,"wave_count":2},"fallback_work":600087584,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1187.0812377929688,"n":16,"spread_pct":55.261469415175654},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.896690836176276,"wave_count":2},"fallback_work":593132480,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":1186.5930786132812,"n":16,"spread_pct":57.7015618579177},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.022302106022835,"wave_count":2},"fallback_work":588932800,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":1174.6217041015625,"n":16,"spread_pct":2.2562845148054547},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.000657936558127,"wave_count":2},"fallback_work":589112800,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":1177.25341796875,"n":16,"spread_pct":59.38485451148168}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.103236133232713,"wave_count":2},"fallback_work":601877920,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1392.6080322265625,"n":16,"spread_pct":113.63748762408241},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":7.610392989590764,"wave_count":3},"fallback_work":1178694208,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":1546.89892578125,"n":16,"spread_pct":100.8343699860703},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":10.222716955468059,"wave_count":1},"fallback_work":1761907424,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":1539.7321166992188,"n":16,"spread_pct":50.25819389890288},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.6924779415130615,"wave_count":2},"fallback_work":599843616,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1399.1143798828125,"n":16,"spread_pct":59.68206729546497},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":7.143658120185137,"wave_count":1},"fallback_work":2345186624,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1832.7799682617188,"n":16,"spread_pct":3.3551320161100207},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.840014036744833,"wave_count":2},"fallback_work":597875296,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1414.7300415039062,"n":16,"spread_pct":15.016400450115553},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.706118954345584,"wave_count":2},"fallback_work":598055296,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1420.310791015625,"n":16,"spread_pct":52.87699741112862}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.881520824506879,"wave_count":2},"fallback_work":587444096,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1418.3980102539062,"n":16,"spread_pct":169.33757097380084},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":11.491035111248493,"wave_count":3},"fallback_work":1168066176,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1518.0054321289062,"n":16,"spread_pct":102.0646093014109}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.940313985571265,"wave_count":2},"fallback_work":597516192,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1341.53369140625,"n":16,"spread_pct":54.771641890458135}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":8192,"summary_ms":7.618672912940383,"wave_count":1},"fallback_work":1020227840,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":1432.4102172851562,"n":16,"spread_pct":7.539739572975971},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":4.211623920127749,"wave_count":2},"fallback_work":796904992,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":1454.1314086914062,"n":16,"spread_pct":3.7159411897564727},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":4.009237978607416,"wave_count":2},"fallback_work":729276256,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1444.3860473632812,"n":16,"spread_pct":114.28784523660669},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.511914914473891,"wave_count":2},"fallback_work":1082754208,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":1428.2891235351562,"n":16,"spread_pct":53.98956433181379},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":5.071334773674607,"wave_count":1},"fallback_work":1355071392,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":1505.313720703125,"n":16,"spread_pct":2.9190376214468112},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":4.828525008633733,"wave_count":2},"fallback_work":1735810464,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":1569.444091796875,"n":16,"spread_pct":114.6468515004333},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":9.74812195636332,"wave_count":1},"fallback_work":2272243872,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":1713.4210815429688,"n":16,"spread_pct":2.4360400647020266},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":6.436916999518871,"wave_count":3},"fallback_work":770069632,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1540.6324462890625,"n":16,"spread_pct":49.907236915135975},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":11.441763024777174,"wave_count":2},"fallback_work":716071680,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1421.0819702148438,"n":16,"spread_pct":5.193581739678212}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7680,"summary_ms":8.151851827278733,"wave_count":2},"fallback_work":827808512,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":1467.453125,"n":16,"spread_pct":4.687774511004397},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.618328159675002,"wave_count":4},"fallback_work":934639136,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":1539.3921508789062,"n":16,"spread_pct":2.737643321241354},{"cp_plan":{"max_rank_tokens":12288,"summary_ms":11.380682000890374,"wave_count":1},"fallback_work":1488344384,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":1511.4816284179688,"n":16,"spread_pct":51.6065040079163},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.030542215332389,"wave_count":2},"fallback_work":604584192,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":1382.5917358398438,"n":16,"spread_pct":4.793473604962445}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7168,"summary_ms":7.267562905326486,"wave_count":4},"fallback_work":645115200,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1605.7691650390625,"n":16,"spread_pct":46.305036747609805},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":5.241688108071685,"wave_count":2},"fallback_work":1232363360,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":1493.4180297851562,"n":16,"spread_pct":98.60375693947739},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":8.771643042564392,"wave_count":2},"fallback_work":1808293792,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":1650.3270263671875,"n":16,"spread_pct":3.285207958087734},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.3292039558291435,"wave_count":1},"fallback_work":2407352096,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":1867.9608154296875,"n":16,"spread_pct":79.76332370190562},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":4.171384032815695,"wave_count":2},"fallback_work":636732896,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1395.2506713867188,"n":16,"spread_pct":2.4570220186216085},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.69262502156198,"wave_count":2},"fallback_work":628023136,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1409.8743896484375,"n":16,"spread_pct":5.728845905920773},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":17.768254037946463,"wave_count":3},"fallback_work":3006082720,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2141.9039306640625,"n":16,"spread_pct":35.54984855364413},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":4.1913369204849005,"wave_count":2},"fallback_work":642441024,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1405.169189453125,"n":16,"spread_pct":58.81864412109793},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.79404185898602,"wave_count":2},"fallback_work":627415808,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1401.3402099609375,"n":16,"spread_pct":5.069060998473925},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.776236834004521,"wave_count":2},"fallback_work":627792416,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1407.9188842773438,"n":16,"spread_pct":164.87921474094352}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":4096,"summary_ms":4.360634833574295,"wave_count":1},"fallback_work":449134048,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1261.7631225585938,"n":16,"spread_pct":5.32505459971488},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.77872702665627,"wave_count":2},"fallback_work":320190688,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":1359.8742065429688,"n":16,"spread_pct":56.63256116568902},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.805382177233696,"wave_count":2},"fallback_work":322159008,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":1359.4641723632812,"n":16,"spread_pct":56.166771746878496},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.8421940989792347,"wave_count":2},"fallback_work":332804768,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":1357.8593139648438,"n":16,"spread_pct":4.52974529296247},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.808459034189582,"wave_count":2},"fallback_work":337085920,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":1357.6395263671875,"n":16,"spread_pct":60.72897663980677},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.8643311709165573,"wave_count":2},"fallback_work":339251072,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":1358.8418579101562,"n":16,"spread_pct":54.81222992225084},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.7299380637705326,"wave_count":1},"fallback_work":350684160,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":1248.70654296875,"n":16,"spread_pct":64.13547651456527},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.717401996254921,"wave_count":2},"fallback_work":363035424,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":1374.16845703125,"n":16,"spread_pct":60.405644820003005},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.743155160918832,"wave_count":2},"fallback_work":366775232,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":1371.7094116210938,"n":16,"spread_pct":61.169835286301065},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.705905983224511,"wave_count":1},"fallback_work":381062496,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":1250.5850219726562,"n":16,"spread_pct":60.632257803171704},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.1431409772485495,"wave_count":2},"fallback_work":302325216,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":1359.9611206054688,"n":16,"spread_pct":66.35238945779619},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.83273814804852,"wave_count":1},"fallback_work":384605472,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":1250.4641723632812,"n":16,"spread_pct":65.56685371472166},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.9800889790058136,"wave_count":2},"fallback_work":303670720,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":1364.5941162109375,"n":16,"spread_pct":55.30552373676675},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.9109059143811464,"wave_count":2},"fallback_work":306657088,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":1360.1139526367188,"n":16,"spread_pct":59.85891200396767},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.802647115662694,"wave_count":2},"fallback_work":398285184,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":1378.2706298828125,"n":16,"spread_pct":61.12816876547112},{"cp_plan":{"max_rank_tokens":4608,"summary_ms":4.3303389102220535,"wave_count":3},"fallback_work":660876128,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":1437.747314453125,"n":16,"spread_pct":58.140693525383625},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.58276504650712,"wave_count":1},"fallback_work":869075008,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":1359.0450439453125,"n":16,"spread_pct":58.62870630326309},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.788933089002967,"wave_count":1},"fallback_work":1089231712,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":1365.6427612304688,"n":16,"spread_pct":7.760714997136721},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":5.315530113875866,"wave_count":3},"fallback_work":1316375840,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":1470.0230712890625,"n":16,"spread_pct":67.3677544621604},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":5.7575879618525505,"wave_count":1},"fallback_work":1546570752,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":1432.4929809570312,"n":16,"spread_pct":5.659701665511644},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.122489063069224,"wave_count":1},"fallback_work":1774207072,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":1494.192626953125,"n":16,"spread_pct":5.868464055090893},{"cp_plan":{"max_rank_tokens":12288,"summary_ms":5.525666056200862,"wave_count":3},"fallback_work":2010208640,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":1541.3283081054688,"n":16,"spread_pct":46.37284663762181},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":10.36056294105947,"wave_count":3},"fallback_work":2243454560,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":1600.2733154296875,"n":16,"spread_pct":4.571919515396502},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.225320976227522,"wave_count":3},"fallback_work":2479849792,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":1633.66455078125,"n":16,"spread_pct":4.148328698092652},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":6.755694048479199,"wave_count":1},"fallback_work":2726381760,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":1692.3927001953125,"n":16,"spread_pct":43.09703614686037},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.869632026180625,"wave_count":1},"fallback_work":417722400,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":1256.009033203125,"n":16,"spread_pct":5.324375225114115},{"cp_plan":{"max_rank_tokens":16896,"summary_ms":7.453965954482555,"wave_count":1},"fallback_work":2978671232,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":1736.2058715820312,"n":16,"spread_pct":3.424903686799939},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.895496018230915,"wave_count":1},"fallback_work":312185216,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":1233.3832397460938,"n":16,"spread_pct":62.06265110847813},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.227854078635573,"wave_count":2},"fallback_work":315745024,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":1365.4966430664062,"n":16,"spread_pct":60.97750266769657},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.853052156046033,"wave_count":2},"fallback_work":316630880,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":1361.1826171875,"n":16,"spread_pct":56.61768187560937},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.8110660389065742,"wave_count":2},"fallback_work":317172224,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":1361.0877075195312,"n":16,"spread_pct":62.09035930439831},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.8565630093216896,"wave_count":2},"fallback_work":316040384,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":1367.7628784179688,"n":16,"spread_pct":64.0665823360856},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":7.8333348501473665,"wave_count":1},"fallback_work":3242935136,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":1885.78369140625,"n":16,"spread_pct":1.0725359488462838},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.052706994116306,"wave_count":1},"fallback_work":314545856,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":1238.6464233398438,"n":16,"spread_pct":65.28643490043329},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":6.154031027108431,"wave_count":2},"fallback_work":307000032,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":1351.9379272460938,"n":16,"spread_pct":58.43189962903639},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.346499918028712,"wave_count":2},"fallback_work":301225920,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":1361.46337890625,"n":16,"spread_pct":111.2014606152832},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":4.027773160487413,"wave_count":1},"fallback_work":428449856,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1246.4320068359375,"n":16,"spread_pct":5.165094780795691},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.970182966440916,"wave_count":2},"fallback_work":402303520,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1380.6802978515625,"n":16,"spread_pct":54.06835380790631},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.951180027797818,"wave_count":1},"fallback_work":385457440,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":1242.7706298828125,"n":16,"spread_pct":7.772864487651581},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.9760421495884657,"wave_count":2},"fallback_work":373679840,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":1370.09765625,"n":16,"spread_pct":114.37099067698755},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":4.006123868748546,"wave_count":2},"fallback_work":363427744,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":1363.5037841796875,"n":16,"spread_pct":61.46006968955551},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":4.14245598949492,"wave_count":2},"fallback_work":349091328,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":1367.3126831054688,"n":16,"spread_pct":57.879495051591476},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.0605319663882256,"wave_count":2},"fallback_work":340659872,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":1362.1845092773438,"n":16,"spread_pct":63.597581704228304},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.066146910190582,"wave_count":2},"fallback_work":327701280,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":1357.5723876953125,"n":16,"spread_pct":58.70055055945686}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.599706968292594,"wave_count":2},"fallback_work":607076576,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1400.0899047851562,"n":16,"spread_pct":58.25459468218267},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":6.865331903100014,"wave_count":3},"fallback_work":1161749152,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":1492.4749755859375,"n":16,"spread_pct":52.45270565103009},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":7.879395037889481,"wave_count":2},"fallback_work":1719718720,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":1609.1224975585938,"n":16,"spread_pct":44.52700273855922},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":6.125960033386946,"wave_count":1},"fallback_work":2273751648,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":1786.2420654296875,"n":16,"spread_pct":41.474205890222},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":12.295401887968183,"wave_count":3},"fallback_work":2826062240,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":2007.685546875,"n":16,"spread_pct":38.07700632822114},{"cp_plan":{"max_rank_tokens":19456,"summary_ms":8.176533039659262,"wave_count":1},"fallback_work":3383588768,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":2324.2235107421875,"n":16,"spread_pct":65.70930868366393},{"cp_plan":{"max_rank_tokens":21504,"summary_ms":9.642038028687239,"wave_count":1},"fallback_work":3947807808,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":2555.8017578125,"n":16,"spread_pct":28.608173957765366},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.9826228749006987,"wave_count":2},"fallback_work":589048160,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":1398.8273315429688,"n":16,"spread_pct":3.8765031762488347},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.8094590418040752,"wave_count":2},"fallback_work":600711968,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1385.0548095703125,"n":16,"spread_pct":4.325674583648743},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.961716080084443,"wave_count":2},"fallback_work":589261824,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1394.8580322265625,"n":16,"spread_pct":4.175850949493015},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.072804003953934,"wave_count":2},"fallback_work":588785120,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":1398.2265625,"n":16,"spread_pct":2.9910412464449943},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.9334429893642664,"wave_count":2},"fallback_work":590607328,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1401.5004272460938,"n":16,"spread_pct":58.89094255541818},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.913413897156715,"wave_count":2},"fallback_work":592674176,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1400.7163696289062,"n":16,"spread_pct":56.676460849391866},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.933008225634694,"wave_count":2},"fallback_work":594807008,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1419.498779296875,"n":16,"spread_pct":53.95386039203222},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.142670892179012,"wave_count":2},"fallback_work":588605344,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":1396.3469848632812,"n":16,"spread_pct":4.14401784021671},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.79342695325613,"wave_count":2},"fallback_work":594413344,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1401.5350341796875,"n":16,"spread_pct":54.43146835635976},{"cp_plan":{"max_rank_tokens":24576,"summary_ms":10.510599007830024,"wave_count":1},"fallback_work":4512092832,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2896.2484130859375,"n":16,"spread_pct":50.01125555280612},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.23600897192955,"wave_count":2},"fallback_work":604254944,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1390.7575073242188,"n":16,"spread_pct":56.05630713470903},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.128149012103677,"wave_count":2},"fallback_work":604188736,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1392.9027709960938,"n":16,"spread_pct":56.71145475859656},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.013679852709174,"wave_count":2},"fallback_work":600087584,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1412.2848510742188,"n":16,"spread_pct":54.07706295141834},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.957892000675201,"wave_count":2},"fallback_work":593132480,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":1406.70458984375,"n":16,"spread_pct":6.262524163195855},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.093857016414404,"wave_count":2},"fallback_work":588932800,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":1399.3284912109375,"n":16,"spread_pct":3.0570606245200445},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.356838857755065,"wave_count":2},"fallback_work":589112800,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":1399.2219848632812,"n":16,"spread_pct":6.585045496256835}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12204,"summary_ms":0.00023655593395233154,"wave_count":0},"fallback_work":602166784,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1091.28369140625,"n":8,"spread_pct":79.82794014539034},{"cp_plan":{"max_rank_tokens":23926,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":1179826944,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":1659.788330078125,"n":8,"spread_pct":4.338983654171594},{"cp_plan":{"max_rank_tokens":35778,"summary_ms":0.00021420419216156006,"wave_count":0},"fallback_work":1763893504,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":2345.1351318359375,"n":8,"spread_pct":1.7448021734238626},{"cp_plan":{"max_rank_tokens":12154,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":599932160,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1086.2240600585938,"n":8,"spread_pct":5.013706196571892},{"cp_plan":{"max_rank_tokens":47640,"summary_ms":0.0003948807716369629,"wave_count":0},"fallback_work":2348223488,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3040.19482421875,"n":8,"spread_pct":2.4665135210707962},{"cp_plan":{"max_rank_tokens":12114,"summary_ms":0.0003203749656677246,"wave_count":0},"fallback_work":597960960,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1098.7477416992188,"n":8,"spread_pct":3.9821723094189645},{"cp_plan":{"max_rank_tokens":12109,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":597943936,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1066.1591796875,"n":8,"spread_pct":5.269490092484797}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11920,"summary_ms":0.00022351741790771484,"wave_count":0},"fallback_work":587909120,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1005.2802734375,"n":8,"spread_pct":5.7341508266409695},{"cp_plan":{"max_rank_tokens":23728,"summary_ms":0.0003203749656677246,"wave_count":0},"fallback_work":1169577984,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1595.876220703125,"n":8,"spread_pct":1.9806832459575607}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12138,"summary_ms":0.00024028122425079346,"wave_count":0},"fallback_work":598291712,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1014.5884094238281,"n":8,"spread_pct":6.561348236645256}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":20701,"summary_ms":0.0003296881914138794,"wave_count":0},"fallback_work":1021030016,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":1571.3355102539062,"n":8,"spread_pct":11.945257522491664},{"cp_plan":{"max_rank_tokens":16154,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":797183232,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":1301.2117309570312,"n":8,"spread_pct":14.220528463435969},{"cp_plan":{"max_rank_tokens":14771,"summary_ms":0.00019371509552001953,"wave_count":0},"fallback_work":729258368,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1307.2823486328125,"n":8,"spread_pct":56.99951920100526},{"cp_plan":{"max_rank_tokens":21963,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":1083450752,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":1557.6173706054688,"n":8,"spread_pct":5.26105737526755},{"cp_plan":{"max_rank_tokens":27497,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":1356166272,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":1788.764892578125,"n":8,"spread_pct":5.409810938801091},{"cp_plan":{"max_rank_tokens":35243,"summary_ms":0.00021792948246002197,"wave_count":0},"fallback_work":1737659776,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":2207.7598876953125,"n":8,"spread_pct":3.4992872098807646},{"cp_plan":{"max_rank_tokens":46153,"summary_ms":0.00033155083656311035,"wave_count":0},"fallback_work":2275075200,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":2856.7490234375,"n":8,"spread_pct":2.8158031744317102},{"cp_plan":{"max_rank_tokens":15600,"summary_ms":0.0003241002559661865,"wave_count":0},"fallback_work":770111488,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1298.3952026367188,"n":8,"spread_pct":16.667858323265364},{"cp_plan":{"max_rank_tokens":14494,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":715837184,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1245.2150268554688,"n":8,"spread_pct":15.198060662194044}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16800,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":828526592,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":1282.1331176757812,"n":8,"spread_pct":3.4594866943130707},{"cp_plan":{"max_rank_tokens":18971,"summary_ms":0.0002253800630569458,"wave_count":0},"fallback_work":935513472,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":1320.79638671875,"n":8,"spread_pct":64.17938174285977},{"cp_plan":{"max_rank_tokens":30232,"summary_ms":0.00032223761081695557,"wave_count":0},"fallback_work":1490226176,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":1917.1083984375,"n":8,"spread_pct":2.132492986925398},{"cp_plan":{"max_rank_tokens":12255,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":604778368,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":1036.4326782226562,"n":8,"spread_pct":2.576280792440312}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13078,"summary_ms":0.0002253800630569458,"wave_count":0},"fallback_work":645368576,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1135.1364135742188,"n":8,"spread_pct":5.47450206540193},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.00023655593395233154,"wave_count":0},"fallback_work":1233476096,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":1739.619384765625,"n":8,"spread_pct":49.065354324851896},{"cp_plan":{"max_rank_tokens":36716,"summary_ms":0.0002384185791015625,"wave_count":0},"fallback_work":1810249216,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":2421.90625,"n":8,"spread_pct":33.744625069353944},{"cp_plan":{"max_rank_tokens":48890,"summary_ms":0.0002253800630569458,"wave_count":0},"fallback_work":2410183936,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":3164.17333984375,"n":8,"spread_pct":1.4631894839486332},{"cp_plan":{"max_rank_tokens":12899,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":636776832,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1111.9839477539062,"n":8,"spread_pct":3.369024595272616},{"cp_plan":{"max_rank_tokens":12722,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":628054272,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1110.6533203125,"n":8,"spread_pct":5.194520574443709},{"cp_plan":{"max_rank_tokens":61066,"summary_ms":0.00040046870708465576,"wave_count":0},"fallback_work":3009987840,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3925.3829345703125,"n":8,"spread_pct":1.4449849187123682},{"cp_plan":{"max_rank_tokens":13015,"summary_ms":0.0002980232238769531,"wave_count":0},"fallback_work":642493312,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1121.4813842773438,"n":8,"spread_pct":2.0673492086876974},{"cp_plan":{"max_rank_tokens":12701,"summary_ms":0.00024028122425079346,"wave_count":0},"fallback_work":627248768,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1102.4867553710938,"n":8,"spread_pct":425.5805426552843},{"cp_plan":{"max_rank_tokens":12700,"summary_ms":0.00030547380447387695,"wave_count":0},"fallback_work":627428864,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1105.2852172851562,"n":8,"spread_pct":1287.3618096820735}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9058,"summary_ms":0.0002253800630569458,"wave_count":0},"fallback_work":448311552,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":962.8709106445312,"n":8,"spread_pct":216.1072390110421},{"cp_plan":{"max_rank_tokens":6403,"summary_ms":0.00018067657947540283,"wave_count":0},"fallback_work":318390656,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":930.64453125,"n":8,"spread_pct":77.05409609855401},{"cp_plan":{"max_rank_tokens":6443,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":320361856,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":932.4680480957031,"n":8,"spread_pct":2.2083047115551495},{"cp_plan":{"max_rank_tokens":6668,"summary_ms":0.00030174851417541504,"wave_count":0},"fallback_work":331220480,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":932.803955078125,"n":8,"spread_pct":1.4829556384650717},{"cp_plan":{"max_rank_tokens":6755,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":335507840,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":932.2044372558594,"n":8,"spread_pct":5.331227252899065},{"cp_plan":{"max_rank_tokens":6799,"summary_ms":0.00021047890186309814,"wave_count":0},"fallback_work":337676160,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":933.6268310546875,"n":8,"spread_pct":3.8173725800249496},{"cp_plan":{"max_rank_tokens":7040,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":349323264,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":938.4671630859375,"n":8,"spread_pct":3.14704634788876},{"cp_plan":{"max_rank_tokens":7291,"summary_ms":0.00021792948246002197,"wave_count":0},"fallback_work":361692544,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":936.87353515625,"n":8,"spread_pct":5.310071782342055},{"cp_plan":{"max_rank_tokens":7367,"summary_ms":0.00034086406230926514,"wave_count":0},"fallback_work":365437824,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":936.6319274902344,"n":8,"spread_pct":11.581303065289458},{"cp_plan":{"max_rank_tokens":7666,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":379943168,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":936.7078247070312,"n":8,"spread_pct":79.27107808088618},{"cp_plan":{"max_rank_tokens":5988,"summary_ms":0.00017881393432617188,"wave_count":0},"fallback_work":299315712,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":929.1785278320312,"n":8,"spread_pct":4.455840155662951},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":383491328,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":937.6515808105469,"n":8,"spread_pct":4.69033538750424},{"cp_plan":{"max_rank_tokens":6024,"summary_ms":0.00019371509552001953,"wave_count":0},"fallback_work":300860416,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":933.3278503417969,"n":8,"spread_pct":3.5735530978079626},{"cp_plan":{"max_rank_tokens":6102,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":304245504,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":935.1719665527344,"n":8,"spread_pct":4.069591431717727},{"cp_plan":{"max_rank_tokens":8016,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":397191168,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":942.3738403320312,"n":8,"spread_pct":2.362781562751925},{"cp_plan":{"max_rank_tokens":13361,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":660363392,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":1034.7266845703125,"n":8,"spread_pct":2.7183450302077006},{"cp_plan":{"max_rank_tokens":17592,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":868867072,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":1182.8215942382812,"n":8,"spread_pct":2.064795881430279},{"cp_plan":{"max_rank_tokens":22066,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":1089345792,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":1362.0943603515625,"n":8,"spread_pct":1.538382673238888},{"cp_plan":{"max_rank_tokens":26682,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":1316822272,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":1580.3164672851562,"n":8,"spread_pct":1.1748768582304006},{"cp_plan":{"max_rank_tokens":31360,"summary_ms":0.00025704503059387207,"wave_count":0},"fallback_work":1547354112,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":1776.1550903320312,"n":8,"spread_pct":40.18973540855363},{"cp_plan":{"max_rank_tokens":35986,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":1775323392,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":2019.6679077148438,"n":8,"spread_pct":1.8096937891563734},{"cp_plan":{"max_rank_tokens":40782,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":2011670272,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":2255.1463623046875,"n":8,"spread_pct":1.0558099018201585},{"cp_plan":{"max_rank_tokens":45522,"summary_ms":0.00023469328880310059,"wave_count":0},"fallback_work":2245257472,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":2488.9451904296875,"n":8,"spread_pct":0.5249678522297745},{"cp_plan":{"max_rank_tokens":50326,"summary_ms":0.0001862645149230957,"wave_count":0},"fallback_work":2481998592,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":2736.805419921875,"n":8,"spread_pct":0.2578868514637213},{"cp_plan":{"max_rank_tokens":55336,"summary_ms":0.00024028122425079346,"wave_count":0},"fallback_work":2728891392,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":2972.673095703125,"n":8,"spread_pct":0.8674556624145281},{"cp_plan":{"max_rank_tokens":8411,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":416656768,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":944.2240905761719,"n":8,"spread_pct":4.0162654993877736},{"cp_plan":{"max_rank_tokens":60463,"summary_ms":0.0002868473529815674,"wave_count":0},"fallback_work":2981549952,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":3235.162353515625,"n":8,"spread_pct":0.7300230733385027},{"cp_plan":{"max_rank_tokens":6223,"summary_ms":0.00018440186977386475,"wave_count":0},"fallback_work":309979008,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":931.7188415527344,"n":8,"spread_pct":5.115713064315257},{"cp_plan":{"max_rank_tokens":6304,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":313741312,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":935.8268737792969,"n":8,"spread_pct":4.671455486557634},{"cp_plan":{"max_rank_tokens":6322,"summary_ms":0.00020116567611694336,"wave_count":0},"fallback_work":314628352,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":939.6921691894531,"n":8,"spread_pct":4.59089265650408},{"cp_plan":{"max_rank_tokens":6333,"summary_ms":0.0002868473529815674,"wave_count":0},"fallback_work":315170432,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":933.5188293457031,"n":8,"spread_pct":4.471135481093295},{"cp_plan":{"max_rank_tokens":6310,"summary_ms":0.00018440186977386475,"wave_count":0},"fallback_work":314036992,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":934.3454895019531,"n":8,"spread_pct":7.279441864165795},{"cp_plan":{"max_rank_tokens":65842,"summary_ms":0.0005383044481277466,"wave_count":0},"fallback_work":3246397696,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":3529.9334716796875,"n":8,"spread_pct":0.8330885252953992},{"cp_plan":{"max_rank_tokens":6245,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":311751296,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":930.9087524414062,"n":8,"spread_pct":4.171614375522595},{"cp_plan":{"max_rank_tokens":6083,"summary_ms":0.00018440186977386475,"wave_count":0},"fallback_work":303997312,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":928.4493713378906,"n":8,"spread_pct":3.591994840555284},{"cp_plan":{"max_rank_tokens":5957,"summary_ms":0.00021047890186309814,"wave_count":0},"fallback_work":298017408,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":937.4607543945312,"n":8,"spread_pct":77.99726902630103},{"cp_plan":{"max_rank_tokens":8629,"summary_ms":0.00023469328880310059,"wave_count":0},"fallback_work":427399808,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":948.9611511230469,"n":8,"spread_pct":438.4779768712311},{"cp_plan":{"max_rank_tokens":8089,"summary_ms":0.00024959444999694824,"wave_count":0},"fallback_work":401017984,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":947.8599853515625,"n":8,"spread_pct":2.8861658838030624},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.00021792948246002197,"wave_count":0},"fallback_work":383950080,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":945.8524169921875,"n":8,"spread_pct":6.089142268046027},{"cp_plan":{"max_rank_tokens":7490,"summary_ms":0.00021047890186309814,"wave_count":0},"fallback_work":371958016,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":943.7531127929688,"n":8,"spread_pct":76.82553690873458},{"cp_plan":{"max_rank_tokens":7273,"summary_ms":0.0001825392246246338,"wave_count":0},"fallback_work":361493632,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":938.5434265136719,"n":8,"spread_pct":1.9014774279056796},{"cp_plan":{"max_rank_tokens":6973,"summary_ms":0.0001825392246246338,"wave_count":0},"fallback_work":346939008,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":936.7444152832031,"n":8,"spread_pct":3.323622707507004},{"cp_plan":{"max_rank_tokens":6793,"summary_ms":0.00030919909477233887,"wave_count":0},"fallback_work":338297984,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":939.2216796875,"n":8,"spread_pct":4.092503818508302},{"cp_plan":{"max_rank_tokens":6521,"summary_ms":0.00020116567611694336,"wave_count":0},"fallback_work":325123200,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":933.7952880859375,"n":8,"spread_pct":78.663353400044}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12291,"summary_ms":0.00023283064365386963,"wave_count":0},"fallback_work":606978432,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1107.7516479492188,"n":8,"spread_pct":3.3206491695622162},{"cp_plan":{"max_rank_tokens":23563,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":1162462592,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":1659.0928344726562,"n":8,"spread_pct":1.6253425861720177},{"cp_plan":{"max_rank_tokens":34902,"summary_ms":0.00017881393432617188,"wave_count":0},"fallback_work":1721248512,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":2313.1895751953125,"n":8,"spread_pct":1.0273435671552544},{"cp_plan":{"max_rank_tokens":46161,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":2276092032,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":2987.5535888671875,"n":8,"spread_pct":23.727280977202007},{"cp_plan":{"max_rank_tokens":57385,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":2829210752,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":3668.4005126953125,"n":8,"spread_pct":23.337728116868032},{"cp_plan":{"max_rank_tokens":68715,"summary_ms":0.00023096799850463867,"wave_count":0},"fallback_work":3387553152,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":4349.682373046875,"n":8,"spread_pct":0.41266707524385393},{"cp_plan":{"max_rank_tokens":80181,"summary_ms":0.00017695128917694092,"wave_count":0},"fallback_work":3952597632,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":5071.864501953125,"n":8,"spread_pct":0.7209079924832331},{"cp_plan":{"max_rank_tokens":11890,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":588134656,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":1081.556640625,"n":8,"spread_pct":6.892113779148616},{"cp_plan":{"max_rank_tokens":12153,"summary_ms":0.00017881393432617188,"wave_count":0},"fallback_work":600407168,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1094.9481811523438,"n":8,"spread_pct":2.9131814361176103},{"cp_plan":{"max_rank_tokens":11903,"summary_ms":0.0002253800630569458,"wave_count":0},"fallback_work":588545920,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1070.1807861328125,"n":8,"spread_pct":2.7114992823592226},{"cp_plan":{"max_rank_tokens":11876,"summary_ms":0.0001862645149230957,"wave_count":0},"fallback_work":587674112,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":1079.5471801757812,"n":8,"spread_pct":17.561639280068793},{"cp_plan":{"max_rank_tokens":11939,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":590090624,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1079.3699951171875,"n":8,"spread_pct":2.832756874166709},{"cp_plan":{"max_rank_tokens":11981,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":592160384,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1078.0243530273438,"n":8,"spread_pct":3.6223490292996665},{"cp_plan":{"max_rank_tokens":12033,"summary_ms":0.00021792948246002197,"wave_count":0},"fallback_work":594493568,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1087.3721923828125,"n":8,"spread_pct":5.367638729991728},{"cp_plan":{"max_rank_tokens":11881,"summary_ms":0.00025704503059387207,"wave_count":0},"fallback_work":587691136,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":1077.8258666992188,"n":8,"spread_pct":2.1741879356359006},{"cp_plan":{"max_rank_tokens":12025,"summary_ms":0.0001825392246246338,"wave_count":0},"fallback_work":594099328,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1086.032470703125,"n":8,"spread_pct":1.8162642700537204},{"cp_plan":{"max_rank_tokens":91657,"summary_ms":0.0003781169652938843,"wave_count":0},"fallback_work":4517905536,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5772.40869140625,"n":8,"spread_pct":2.9255551543225775},{"cp_plan":{"max_rank_tokens":12225,"summary_ms":0.00024214386940002441,"wave_count":0},"fallback_work":603955328,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1095.3062744140625,"n":8,"spread_pct":563.5883495682316},{"cp_plan":{"max_rank_tokens":12215,"summary_ms":0.00023469328880310059,"wave_count":0},"fallback_work":603691904,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1095.9091796875,"n":8,"spread_pct":2.269949679694817},{"cp_plan":{"max_rank_tokens":12123,"summary_ms":0.00036694109439849854,"wave_count":0},"fallback_work":599387520,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1087.28515625,"n":8,"spread_pct":3.2097082757010185},{"cp_plan":{"max_rank_tokens":11973,"summary_ms":0.00017508864402770996,"wave_count":0},"fallback_work":592224896,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":1089.4029541015625,"n":8,"spread_pct":83.32336252935634},{"cp_plan":{"max_rank_tokens":11879,"summary_ms":0.0002253800630569458,"wave_count":0},"fallback_work":587821952,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":1071.4290161132812,"n":8,"spread_pct":3.8651741770517334},{"cp_plan":{"max_rank_tokens":11874,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":587804928,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":1078.8241577148438,"n":8,"spread_pct":144.37866521682184}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12204,"summary_ms":0.0005159527063369751,"wave_count":0},"fallback_work":602166784,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1157.4345092773438,"n":8,"spread_pct":64.35090487695592},{"cp_plan":{"max_rank_tokens":23926,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":1179826944,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":1630.3766479492188,"n":8,"spread_pct":1.4635619502793697},{"cp_plan":{"max_rank_tokens":35778,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":1763893504,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":2245.77734375,"n":8,"spread_pct":0.6854443843393591},{"cp_plan":{"max_rank_tokens":12154,"summary_ms":0.00018067657947540283,"wave_count":0},"fallback_work":599932160,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1146.3966674804688,"n":8,"spread_pct":1.334418524245067},{"cp_plan":{"max_rank_tokens":47640,"summary_ms":0.0003948807716369629,"wave_count":0},"fallback_work":2348223488,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2890.6705322265625,"n":8,"spread_pct":1.0941381708810076},{"cp_plan":{"max_rank_tokens":12114,"summary_ms":0.00028312206268310547,"wave_count":0},"fallback_work":597960960,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1149.7155151367188,"n":8,"spread_pct":3.5850835425012524},{"cp_plan":{"max_rank_tokens":12109,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":597943936,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1152.4236450195312,"n":8,"spread_pct":64.22926503665609}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11920,"summary_ms":0.0003557652235031128,"wave_count":0},"fallback_work":587909120,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1112.2918090820312,"n":8,"spread_pct":8.472189839335826},{"cp_plan":{"max_rank_tokens":23728,"summary_ms":0.00040978193283081055,"wave_count":0},"fallback_work":1169577984,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1585.877685546875,"n":8,"spread_pct":1.2724771793271143}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12138,"summary_ms":0.00034831464290618896,"wave_count":0},"fallback_work":598291712,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1100.218017578125,"n":8,"spread_pct":77.92070241760057}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":20701,"summary_ms":0.00040978193283081055,"wave_count":0},"fallback_work":1021030016,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":1519.9835815429688,"n":8,"spread_pct":1.335945505617684},{"cp_plan":{"max_rank_tokens":16154,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":797183232,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":1297.8767700195312,"n":8,"spread_pct":57.533880667197344},{"cp_plan":{"max_rank_tokens":14771,"summary_ms":0.0001825392246246338,"wave_count":0},"fallback_work":729258368,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1243.0833740234375,"n":8,"spread_pct":64.25438650328738},{"cp_plan":{"max_rank_tokens":21963,"summary_ms":0.00017695128917694092,"wave_count":0},"fallback_work":1083450752,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":1499.3562622070312,"n":8,"spread_pct":1.4013256747188827},{"cp_plan":{"max_rank_tokens":27497,"summary_ms":0.0001862645149230957,"wave_count":0},"fallback_work":1356166272,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":1711.271728515625,"n":8,"spread_pct":1.18619807289358},{"cp_plan":{"max_rank_tokens":35243,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":1737659776,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":2067.6907958984375,"n":8,"spread_pct":1.1712119173554862},{"cp_plan":{"max_rank_tokens":46153,"summary_ms":0.00038370490074157715,"wave_count":0},"fallback_work":2275075200,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":2647.5390625,"n":8,"spread_pct":0.5269578913200643},{"cp_plan":{"max_rank_tokens":15600,"summary_ms":0.0003688037395477295,"wave_count":0},"fallback_work":770111488,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1274.3818969726562,"n":8,"spread_pct":1.3643352013721803},{"cp_plan":{"max_rank_tokens":14494,"summary_ms":0.0002551823854446411,"wave_count":0},"fallback_work":715837184,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1245.7260131835938,"n":8,"spread_pct":2.045215635981916}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16800,"summary_ms":0.00023655593395233154,"wave_count":0},"fallback_work":828526592,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":1333.9752807617188,"n":8,"spread_pct":1.8566195386306394},{"cp_plan":{"max_rank_tokens":18971,"summary_ms":0.0001825392246246338,"wave_count":0},"fallback_work":935513472,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":1358.3317260742188,"n":8,"spread_pct":1.989871267491704},{"cp_plan":{"max_rank_tokens":30232,"summary_ms":0.00028870999813079834,"wave_count":0},"fallback_work":1490226176,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":1867.9501342773438,"n":8,"spread_pct":0.8568937340740306},{"cp_plan":{"max_rank_tokens":12255,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":604778368,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":1133.0165405273438,"n":8,"spread_pct":3.233695568734576}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13078,"summary_ms":0.0002477318048477173,"wave_count":0},"fallback_work":645368576,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1222.1300659179688,"n":8,"spread_pct":1.0158625075218324},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":1233476096,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":1723.5191650390625,"n":8,"spread_pct":1.9528270596311978},{"cp_plan":{"max_rank_tokens":36716,"summary_ms":0.00018440186977386475,"wave_count":0},"fallback_work":1810249216,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":2333.11328125,"n":8,"spread_pct":1.2939766640938792},{"cp_plan":{"max_rank_tokens":48890,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":2410183936,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":3025.77880859375,"n":8,"spread_pct":0.6519095758371868},{"cp_plan":{"max_rank_tokens":12899,"summary_ms":0.00017881393432617188,"wave_count":0},"fallback_work":636776832,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1210.0114135742188,"n":8,"spread_pct":1.54231862906042},{"cp_plan":{"max_rank_tokens":12722,"summary_ms":0.00018998980522155762,"wave_count":0},"fallback_work":628054272,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1203.8001098632812,"n":8,"spread_pct":3.3535666700011726},{"cp_plan":{"max_rank_tokens":61066,"summary_ms":0.00028870999813079834,"wave_count":0},"fallback_work":3009987840,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3729.1256103515625,"n":8,"spread_pct":19.418529518334104},{"cp_plan":{"max_rank_tokens":13015,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":642493312,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1219.058837890625,"n":8,"spread_pct":2.2294030058488783},{"cp_plan":{"max_rank_tokens":12701,"summary_ms":0.0001862645149230957,"wave_count":0},"fallback_work":627248768,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1200.5917358398438,"n":8,"spread_pct":2.3749986972874857},{"cp_plan":{"max_rank_tokens":12700,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":627428864,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1207.2799682617188,"n":8,"spread_pct":72.57655999787666}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9058,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":448311552,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1100.1052856445312,"n":8,"spread_pct":72.6265491390012},{"cp_plan":{"max_rank_tokens":6403,"summary_ms":0.00017695128917694092,"wave_count":0},"fallback_work":318390656,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":1080.229736328125,"n":8,"spread_pct":5.339881088120316},{"cp_plan":{"max_rank_tokens":6443,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":320361856,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":1081.197021484375,"n":8,"spread_pct":2.1816910736459043},{"cp_plan":{"max_rank_tokens":6668,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":331220480,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":1081.44873046875,"n":8,"spread_pct":2.8033706774450327},{"cp_plan":{"max_rank_tokens":6755,"summary_ms":0.00019744038581848145,"wave_count":0},"fallback_work":335507840,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":1091.6991577148438,"n":8,"spread_pct":69.03549451177959},{"cp_plan":{"max_rank_tokens":6799,"summary_ms":0.00019371509552001953,"wave_count":0},"fallback_work":337676160,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":1085.93701171875,"n":8,"spread_pct":4.527104103913715},{"cp_plan":{"max_rank_tokens":7040,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":349323264,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":1091.8270874023438,"n":8,"spread_pct":3.151847039116482},{"cp_plan":{"max_rank_tokens":7291,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":361692544,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":1093.7402954101562,"n":8,"spread_pct":7.051692032312117},{"cp_plan":{"max_rank_tokens":7367,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":365437824,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":1090.8517456054688,"n":8,"spread_pct":3.9009059441270573},{"cp_plan":{"max_rank_tokens":7666,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":379943168,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":1096.211181640625,"n":8,"spread_pct":67.2124845854674},{"cp_plan":{"max_rank_tokens":5988,"summary_ms":0.00018440186977386475,"wave_count":0},"fallback_work":299315712,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":1083.58154296875,"n":8,"spread_pct":2.7572070701950047},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.00018067657947540283,"wave_count":0},"fallback_work":383491328,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":1102.819580078125,"n":8,"spread_pct":5.425988826137902},{"cp_plan":{"max_rank_tokens":6024,"summary_ms":0.00018067657947540283,"wave_count":0},"fallback_work":300860416,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":1086.4446411132812,"n":8,"spread_pct":67.93043873564217},{"cp_plan":{"max_rank_tokens":6102,"summary_ms":0.00017508864402770996,"wave_count":0},"fallback_work":304245504,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":1083.4292602539062,"n":8,"spread_pct":2.4142093036946566},{"cp_plan":{"max_rank_tokens":8016,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":397191168,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":1092.4373779296875,"n":8,"spread_pct":2.232042539444939},{"cp_plan":{"max_rank_tokens":13361,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":660363392,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":1141.63232421875,"n":8,"spread_pct":1.1704988274481911},{"cp_plan":{"max_rank_tokens":17592,"summary_ms":0.00020116567611694336,"wave_count":0},"fallback_work":868867072,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":1246.3599853515625,"n":8,"spread_pct":58.95131535866015},{"cp_plan":{"max_rank_tokens":22066,"summary_ms":0.00020116567611694336,"wave_count":0},"fallback_work":1089345792,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":1373.1482543945312,"n":8,"spread_pct":1.7294331175138609},{"cp_plan":{"max_rank_tokens":26682,"summary_ms":0.0002216547727584839,"wave_count":0},"fallback_work":1316822272,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":1542.3504638671875,"n":8,"spread_pct":2.4468428211146316},{"cp_plan":{"max_rank_tokens":31360,"summary_ms":0.000173225998878479,"wave_count":0},"fallback_work":1547354112,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":1720.3707275390625,"n":8,"spread_pct":2.2482989584324495},{"cp_plan":{"max_rank_tokens":35986,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":1775323392,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":1934.9660034179688,"n":8,"spread_pct":38.29587732923002},{"cp_plan":{"max_rank_tokens":40782,"summary_ms":0.00023469328880310059,"wave_count":0},"fallback_work":2011670272,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":2139.5911865234375,"n":8,"spread_pct":0.6967210612835316},{"cp_plan":{"max_rank_tokens":45522,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":2245257472,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":2352.8795166015625,"n":8,"spread_pct":0.7544466504925164},{"cp_plan":{"max_rank_tokens":50326,"summary_ms":0.00020116567611694336,"wave_count":0},"fallback_work":2481998592,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":2579.04296875,"n":8,"spread_pct":1.041788151188592},{"cp_plan":{"max_rank_tokens":55336,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":2728891392,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":2797.724609375,"n":8,"spread_pct":0.3938049545005533},{"cp_plan":{"max_rank_tokens":8411,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":416656768,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":1100.9750366210938,"n":8,"spread_pct":71.43160068465147},{"cp_plan":{"max_rank_tokens":60463,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":2981549952,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":3040.6041259765625,"n":8,"spread_pct":24.756577329085435},{"cp_plan":{"max_rank_tokens":6223,"summary_ms":0.0002253800630569458,"wave_count":0},"fallback_work":309979008,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":1085.7362670898438,"n":8,"spread_pct":3.2697830752472514},{"cp_plan":{"max_rank_tokens":6304,"summary_ms":0.00018067657947540283,"wave_count":0},"fallback_work":313741312,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":1079.0405883789062,"n":8,"spread_pct":3.964183270310353},{"cp_plan":{"max_rank_tokens":6322,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":314628352,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":1080.5096435546875,"n":8,"spread_pct":3.3512831390261693},{"cp_plan":{"max_rank_tokens":6333,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":315170432,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":1077.75048828125,"n":8,"spread_pct":2.8030796929911794},{"cp_plan":{"max_rank_tokens":6310,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":314036992,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":1084.3714599609375,"n":8,"spread_pct":72.35683068579903},{"cp_plan":{"max_rank_tokens":65842,"summary_ms":0.00033527612686157227,"wave_count":0},"fallback_work":3246397696,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":3315.3062744140625,"n":8,"spread_pct":21.892677963822585},{"cp_plan":{"max_rank_tokens":6245,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":311751296,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":1086.9913330078125,"n":8,"spread_pct":1.102920243877541},{"cp_plan":{"max_rank_tokens":6083,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":303997312,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":1086.00927734375,"n":8,"spread_pct":76.70642947610928},{"cp_plan":{"max_rank_tokens":5957,"summary_ms":0.00024028122425079346,"wave_count":0},"fallback_work":298017408,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":1081.287109375,"n":8,"spread_pct":67.88572104642363},{"cp_plan":{"max_rank_tokens":8629,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":427399808,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1094.0768432617188,"n":8,"spread_pct":2.6747252704339743},{"cp_plan":{"max_rank_tokens":8089,"summary_ms":0.00018067657947540283,"wave_count":0},"fallback_work":401017984,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1093.3931884765625,"n":8,"spread_pct":1.8026304786706644},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.00017695128917694092,"wave_count":0},"fallback_work":383950080,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":1094.8093872070312,"n":8,"spread_pct":9.449385436215604},{"cp_plan":{"max_rank_tokens":7490,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":371958016,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":1090.0038452148438,"n":8,"spread_pct":2.96918748998733},{"cp_plan":{"max_rank_tokens":7273,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":361493632,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":1089.3214111328125,"n":8,"spread_pct":70.26148621186162},{"cp_plan":{"max_rank_tokens":6973,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":346939008,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":1084.9921264648438,"n":8,"spread_pct":4.66652876934062},{"cp_plan":{"max_rank_tokens":6793,"summary_ms":0.00017695128917694092,"wave_count":0},"fallback_work":338297984,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":1084.0477294921875,"n":8,"spread_pct":8.14263220426644},{"cp_plan":{"max_rank_tokens":6521,"summary_ms":0.00017881393432617188,"wave_count":0},"fallback_work":325123200,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":1085.3455200195312,"n":8,"spread_pct":7.163246196316214}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12291,"summary_ms":0.000260770320892334,"wave_count":0},"fallback_work":606978432,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1218.10986328125,"n":8,"spread_pct":2.7638615474714485},{"cp_plan":{"max_rank_tokens":23563,"summary_ms":0.00017881393432617188,"wave_count":0},"fallback_work":1162462592,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":1664.6648559570312,"n":8,"spread_pct":45.93136382667233},{"cp_plan":{"max_rank_tokens":34902,"summary_ms":0.0004116445779800415,"wave_count":0},"fallback_work":1721248512,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":2234.349365234375,"n":8,"spread_pct":0.632983660760968},{"cp_plan":{"max_rank_tokens":46161,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":2276092032,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":2849.833740234375,"n":8,"spread_pct":0.600526740569347},{"cp_plan":{"max_rank_tokens":57385,"summary_ms":0.00018440186977386475,"wave_count":0},"fallback_work":2829210752,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":3483.8511962890625,"n":8,"spread_pct":0.5462424301523187},{"cp_plan":{"max_rank_tokens":68715,"summary_ms":0.0002477318048477173,"wave_count":0},"fallback_work":3387553152,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":4137.327392578125,"n":8,"spread_pct":1.180362214176113},{"cp_plan":{"max_rank_tokens":80181,"summary_ms":0.00030174851417541504,"wave_count":0},"fallback_work":3952597632,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":4809.531005859375,"n":8,"spread_pct":0.4240643794093952},{"cp_plan":{"max_rank_tokens":11890,"summary_ms":0.00024586915969848633,"wave_count":0},"fallback_work":588134656,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":1187.3673095703125,"n":8,"spread_pct":1.7668298256600015},{"cp_plan":{"max_rank_tokens":12153,"summary_ms":0.0002216547727584839,"wave_count":0},"fallback_work":600407168,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1208.9873046875,"n":8,"spread_pct":70.92009268152015},{"cp_plan":{"max_rank_tokens":11903,"summary_ms":0.00017881393432617188,"wave_count":0},"fallback_work":588545920,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1188.2944946289062,"n":8,"spread_pct":4.171499773203752},{"cp_plan":{"max_rank_tokens":11876,"summary_ms":0.00023283064365386963,"wave_count":0},"fallback_work":587674112,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":1189.34326171875,"n":8,"spread_pct":1.608954029005142},{"cp_plan":{"max_rank_tokens":11939,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":590090624,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1194.168701171875,"n":8,"spread_pct":3.1897761644874643},{"cp_plan":{"max_rank_tokens":11981,"summary_ms":0.00021420419216156006,"wave_count":0},"fallback_work":592160384,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1196.9881591796875,"n":8,"spread_pct":2.2774547975892045},{"cp_plan":{"max_rank_tokens":12033,"summary_ms":0.00024400651454925537,"wave_count":0},"fallback_work":594493568,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1201.333740234375,"n":8,"spread_pct":4.2939335613920315},{"cp_plan":{"max_rank_tokens":11881,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":587691136,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":1193.7215576171875,"n":8,"spread_pct":3.559414813446042},{"cp_plan":{"max_rank_tokens":12025,"summary_ms":0.00025331974029541016,"wave_count":0},"fallback_work":594099328,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1198.8018188476562,"n":8,"spread_pct":4.6057034588929},{"cp_plan":{"max_rank_tokens":91657,"summary_ms":0.0004041939973831177,"wave_count":0},"fallback_work":4517905536,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5477.091064453125,"n":8,"spread_pct":0.10123844591496943},{"cp_plan":{"max_rank_tokens":12225,"summary_ms":0.0002253800630569458,"wave_count":0},"fallback_work":603955328,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1203.3885498046875,"n":8,"spread_pct":2.5345401712429267},{"cp_plan":{"max_rank_tokens":12215,"summary_ms":0.00023283064365386963,"wave_count":0},"fallback_work":603691904,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1200.5005493164062,"n":8,"spread_pct":70.56449892905084},{"cp_plan":{"max_rank_tokens":12123,"summary_ms":0.0002216547727584839,"wave_count":0},"fallback_work":599387520,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1217.0863647460938,"n":8,"spread_pct":72.26819983588375},{"cp_plan":{"max_rank_tokens":11973,"summary_ms":0.0002477318048477173,"wave_count":0},"fallback_work":592224896,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":1198.4862670898438,"n":8,"spread_pct":62.56675812623539},{"cp_plan":{"max_rank_tokens":11879,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":587821952,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":1190.412841796875,"n":8,"spread_pct":1.525934226714857},{"cp_plan":{"max_rank_tokens":11874,"summary_ms":0.0002253800630569458,"wave_count":0},"fallback_work":587804928,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":1194.9945068359375,"n":8,"spread_pct":3.379207805998226}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":38879,"summary_ms":8.357511833310127,"wave_count":1},"fallback_work":3684103680,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":3956.8399658203125,"n":16,"spread_pct":20.157804395344506},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":8.09571985155344,"wave_count":1},"fallback_work":709948768,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1324.8021240234375,"n":16,"spread_pct":5.353697886965173},{"cp_plan":{"max_rank_tokens":55296,"summary_ms":10.346478084102273,"wave_count":1},"fallback_work":5245103104,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":5027.88232421875,"n":16,"spread_pct":3.7431110478354928},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":8.313086116686463,"wave_count":1},"fallback_work":659879424,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1332.7893676757812,"n":16,"spread_pct":62.88487476954213}],"cell":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":38879,"summary_ms":8.266452001407743,"wave_count":1},"fallback_work":3684103680,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":4156.52685546875,"n":16,"spread_pct":17.19475453174113},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":8.070101030170918,"wave_count":1},"fallback_work":709948768,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1203.2108764648438,"n":16,"spread_pct":59.510149615436845},{"cp_plan":{"max_rank_tokens":55296,"summary_ms":10.100596118718386,"wave_count":1},"fallback_work":5245103104,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":5284.571533203125,"n":16,"spread_pct":13.217861085425886},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":8.233478059992194,"wave_count":1},"fallback_work":659879424,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1189.2051391601562,"n":16,"spread_pct":58.355362163201406}],"cell":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19456,"summary_ms":8.155781077221036,"wave_count":1},"fallback_work":3684538624,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":2118.017333984375,"n":16,"spread_pct":66.48307472659182},{"cp_plan":{"max_rank_tokens":4608,"summary_ms":4.7936060000211,"wave_count":3},"fallback_work":711964928,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1379.5363159179688,"n":16,"spread_pct":50.396819831306715},{"cp_plan":{"max_rank_tokens":26880,"summary_ms":11.972186854109168,"wave_count":1},"fallback_work":5244645888,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":2562.5499267578125,"n":16,"spread_pct":27.685436197436893},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":5.038870964199305,"wave_count":1},"fallback_work":662051072,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1263.2723388671875,"n":16,"spread_pct":115.8606560980882}],"cell":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19456,"summary_ms":12.207899242639542,"wave_count":1},"fallback_work":3684538624,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":2226.4124755859375,"n":8,"spread_pct":31.505938640273513},{"cp_plan":{"max_rank_tokens":4608,"summary_ms":4.4931210577487946,"wave_count":3},"fallback_work":711964928,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1235.5985717773438,"n":8,"spread_pct":310.75221273690977},{"cp_plan":{"max_rank_tokens":26880,"summary_ms":9.371688589453697,"wave_count":1},"fallback_work":5244645888,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":2696.5205078125,"n":8,"spread_pct":25.316226024785273},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":4.596708342432976,"wave_count":1},"fallback_work":662051072,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1121.9532470703125,"n":8,"spread_pct":432.8840334843464}],"cell":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19456,"summary_ms":8.257784880697727,"wave_count":1},"fallback_work":3684538624,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":2575.8338623046875,"n":16,"spread_pct":26.30989632784931},{"cp_plan":{"max_rank_tokens":4608,"summary_ms":4.773485939949751,"wave_count":3},"fallback_work":711964928,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1216.7901611328125,"n":16,"spread_pct":69.46405703482513},{"cp_plan":{"max_rank_tokens":26880,"summary_ms":11.411450104787946,"wave_count":1},"fallback_work":5244645888,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":3186.7802734375,"n":16,"spread_pct":0.5641515869238857},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":4.973966861143708,"wave_count":1},"fallback_work":662051072,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1111.6219482421875,"n":16,"spread_pct":58.87180311620159}],"cell":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":74719,"summary_ms":0.0003781169652938843,"wave_count":0},"fallback_work":3686576000,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":3671.0479736328125,"n":8,"spread_pct":0.3486689571869994},{"cp_plan":{"max_rank_tokens":14302,"summary_ms":0.0003241002559661865,"wave_count":0},"fallback_work":709455616,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1108.3286743164062,"n":8,"spread_pct":5.961939124252149},{"cp_plan":{"max_rank_tokens":106432,"summary_ms":0.0005513429641723633,"wave_count":0},"fallback_work":5249163264,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":4814.881591796875,"n":8,"spread_pct":0.5077844238507171},{"cp_plan":{"max_rank_tokens":13279,"summary_ms":0.000260770320892334,"wave_count":0},"fallback_work":659271552,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1106.8004760742188,"n":8,"spread_pct":2.365072685461347}],"cell":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":74719,"summary_ms":0.00046938657760620117,"wave_count":0},"fallback_work":3686576000,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":4024.5643310546875,"n":8,"spread_pct":17.490762826709545},{"cp_plan":{"max_rank_tokens":14302,"summary_ms":0.0003203749656677246,"wave_count":0},"fallback_work":709455616,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1047.9488525390625,"n":8,"spread_pct":3.6039291319293865},{"cp_plan":{"max_rank_tokens":106432,"summary_ms":0.0006128102540969849,"wave_count":0},"fallback_work":5249163264,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":5287.07958984375,"n":8,"spread_pct":0.7139031721549225},{"cp_plan":{"max_rank_tokens":13279,"summary_ms":0.00043213367462158203,"wave_count":0},"fallback_work":659271552,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1008.7687377929688,"n":8,"spread_pct":79.71573967457113}],"cell":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":58351,"summary_ms":0.00039301812648773193,"wave_count":0},"fallback_work":2877863808,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":5205.060302734375,"n":8,"spread_pct":39.52434733946322},{"cp_plan":{"max_rank_tokens":24556,"summary_ms":0.0003986060619354248,"wave_count":0},"fallback_work":1213134336,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":2200.634521484375,"n":8,"spread_pct":0.6288580430959243},{"cp_plan":{"max_rank_tokens":73696,"summary_ms":0.0007562339305877686,"wave_count":0},"fallback_work":3633836032,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":6162.78076171875,"n":8,"spread_pct":11.848130651375147},{"cp_plan":{"max_rank_tokens":21487,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":1062123392,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":2022.0758056640625,"n":8,"spread_pct":1.027640162041963}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":31727,"summary_ms":11.69571909122169,"wave_count":1},"fallback_work":2875652864,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3340.1463623046875,"n":16,"spread_pct":0.9085068644547886},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":9.26804100163281,"wave_count":1},"fallback_work":1212742080,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1569.8656616210938,"n":16,"spread_pct":45.28294140290775},{"cp_plan":{"max_rank_tokens":37888,"summary_ms":9.194779209792614,"wave_count":1},"fallback_work":3630822912,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":3579.0394287109375,"n":16,"spread_pct":21.05522909657606},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":10.787941049784422,"wave_count":1},"fallback_work":1061944064,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1535.8582763671875,"n":16,"spread_pct":334.8039553477096}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":31727,"summary_ms":11.847594985738397,"wave_count":1},"fallback_work":2875652864,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3435.3204345703125,"n":16,"spread_pct":1.0364099822525965},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":9.319168981164694,"wave_count":1},"fallback_work":1212742080,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1519.4044189453125,"n":16,"spread_pct":45.85240525779747},{"cp_plan":{"max_rank_tokens":37888,"summary_ms":9.216872043907642,"wave_count":1},"fallback_work":3630822912,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":3834.969970703125,"n":16,"spread_pct":0.7276466180974746},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":11.03927893564105,"wave_count":1},"fallback_work":1061944064,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1458.209228515625,"n":16,"spread_pct":47.818955017098276}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":15360,"summary_ms":11.358503950759768,"wave_count":1},"fallback_work":2875432064,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1765.2033081054688,"n":16,"spread_pct":2.2474463359629135},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.536182085052133,"wave_count":1},"fallback_work":1213725600,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1359.3042602539062,"n":16,"spread_pct":58.90415184590469},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":9.126226184889674,"wave_count":1},"fallback_work":3630102784,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":1946.425537109375,"n":16,"spread_pct":1.119326931290192},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.311727058142424,"wave_count":1},"fallback_work":1063132288,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1306.0045166015625,"n":16,"spread_pct":60.07349990732595}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":15360,"summary_ms":7.284002378582954,"wave_count":1},"fallback_work":2875432064,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1835.5910034179688,"n":8,"spread_pct":36.64010657862523},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.678111523389816,"wave_count":1},"fallback_work":1213725600,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1229.8797607421875,"n":8,"spread_pct":4.7453667057892295},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":7.036527618765831,"wave_count":1},"fallback_work":3630102784,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":2029.494873046875,"n":8,"spread_pct":68.56637905035974},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.701743230223656,"wave_count":1},"fallback_work":1063132288,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1204.2498168945312,"n":8,"spread_pct":2.255034596561484}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":15360,"summary_ms":7.866173051297665,"wave_count":1},"fallback_work":2875432064,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":2087.386962890625,"n":16,"spread_pct":31.10494264460662},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.404678173363209,"wave_count":1},"fallback_work":1213725600,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1222.5916748046875,"n":16,"spread_pct":4.282165062431911},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":9.09239100292325,"wave_count":1},"fallback_work":3630102784,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":2356.06689453125,"n":16,"spread_pct":31.130797009465883},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.2791801281273365,"wave_count":1},"fallback_work":1063132288,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1168.165283203125,"n":16,"spread_pct":57.43734593154789}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":58351,"summary_ms":0.00033155083656311035,"wave_count":0},"fallback_work":2877863808,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":2973.49267578125,"n":8,"spread_pct":0.8508529055474371},{"cp_plan":{"max_rank_tokens":24556,"summary_ms":0.00018067657947540283,"wave_count":0},"fallback_work":1213134336,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1394.2139282226562,"n":8,"spread_pct":52.79222981454858},{"cp_plan":{"max_rank_tokens":73696,"summary_ms":0.00041350722312927246,"wave_count":0},"fallback_work":3633836032,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":3487.9256591796875,"n":8,"spread_pct":143.81623518662857},{"cp_plan":{"max_rank_tokens":21487,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":1062123392,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1335.1206665039062,"n":8,"spread_pct":56.28677292955575}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":58351,"summary_ms":0.000346451997756958,"wave_count":0},"fallback_work":2877863808,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3250.6767578125,"n":8,"spread_pct":1.3241984166157708},{"cp_plan":{"max_rank_tokens":24556,"summary_ms":0.0002514570951461792,"wave_count":0},"fallback_work":1213134336,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1454.0800170898438,"n":8,"spread_pct":2.4975940921219997},{"cp_plan":{"max_rank_tokens":73696,"summary_ms":0.0004116445779800415,"wave_count":0},"fallback_work":3633836032,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":3805.4874267578125,"n":8,"spread_pct":1.3006762278195296},{"cp_plan":{"max_rank_tokens":21487,"summary_ms":0.00031851232051849365,"wave_count":0},"fallback_work":1062123392,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1371.9304809570312,"n":8,"spread_pct":62.076984803033085}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":37871,"summary_ms":0.0009778887033462524,"wave_count":0},"fallback_work":1868609408,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3239.0380859375,"n":8,"spread_pct":27.790187757686308},{"cp_plan":{"max_rank_tokens":10222,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":506296064,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1121.9547729492188,"n":8,"spread_pct":2.3078203167895825},{"cp_plan":{"max_rank_tokens":53216,"summary_ms":0.0013206154108047485,"wave_count":0},"fallback_work":2624581632,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":4247.752685546875,"n":8,"spread_pct":2.4452957805411235},{"cp_plan":{"max_rank_tokens":9199,"summary_ms":0.0004637986421585083,"wave_count":0},"fallback_work":456112000,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1090.309814453125,"n":8,"spread_pct":320.13665079158574}],"cell":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":21487,"summary_ms":9.028061991557479,"wave_count":1},"fallback_work":1867381504,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":2228.18505859375,"n":16,"spread_pct":1.4090061124316644},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.685344967991114,"wave_count":1},"fallback_work":506460768,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1260.0531616210938,"n":16,"spread_pct":71.99387775396808},{"cp_plan":{"max_rank_tokens":28672,"summary_ms":25.22244700230658,"wave_count":1},"fallback_work":2622551552,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":2600.61376953125,"n":16,"spread_pct":4.4003941377270275},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":8.546791970729828,"wave_count":1},"fallback_work":456391424,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1265.153076171875,"n":16,"spread_pct":70.43991326241054}],"cell":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":21487,"summary_ms":8.720651967450976,"wave_count":1},"fallback_work":1867381504,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":2240.181396484375,"n":16,"spread_pct":29.55594089354554},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.64287805557251,"wave_count":1},"fallback_work":506460768,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1083.9091186523438,"n":16,"spread_pct":2.366481862169823},{"cp_plan":{"max_rank_tokens":28672,"summary_ms":10.341065004467964,"wave_count":1},"fallback_work":2622551552,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":2680.2098388671875,"n":16,"spread_pct":25.067007027921264},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":8.459731005132198,"wave_count":1},"fallback_work":456391424,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1076.9437255859375,"n":16,"spread_pct":64.3172325474821}],"cell":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.718092201277614,"wave_count":1},"fallback_work":1867652224,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1418.781982421875,"n":16,"spread_pct":51.20931848487034},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":4.157983930781484,"wave_count":3},"fallback_work":507526272,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1342.9729614257812,"n":16,"spread_pct":56.94199348287747},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":8.431431837379932,"wave_count":1},"fallback_work":2622322944,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":1535.03125,"n":16,"spread_pct":2.922860449196881},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.303022986277938,"wave_count":1},"fallback_work":457612416,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1223.2373657226562,"n":16,"spread_pct":72.33202509929262}],"cell":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.647772923111916,"wave_count":1},"fallback_work":1867652224,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1431.7545776367188,"n":8,"spread_pct":3.322211691284634},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.754636272788048,"wave_count":3},"fallback_work":507526272,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1312.5079956054688,"n":8,"spread_pct":18.905037362356097},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":6.810208782553673,"wave_count":1},"fallback_work":2622322944,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":1635.4694213867188,"n":8,"spread_pct":43.21472860530362},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.895878791809082,"wave_count":1},"fallback_work":457612416,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1185.6693115234375,"n":8,"spread_pct":5.1382049403258705}],"cell":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.497853038832545,"wave_count":1},"fallback_work":1867652224,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1471.1204833984375,"n":16,"spread_pct":77.61682669899702},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":4.043795866891742,"wave_count":3},"fallback_work":507526272,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1166.6982421875,"n":16,"spread_pct":54.51408681182457},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":7.972029037773609,"wave_count":1},"fallback_work":2622322944,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":1751.3226928710938,"n":16,"spread_pct":3.139807381216576},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.144964041188359,"wave_count":1},"fallback_work":457612416,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1016.3144226074219,"n":16,"spread_pct":7.155305123256025}],"cell":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":37871,"summary_ms":0.0004675239324569702,"wave_count":0},"fallback_work":1868609408,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1910.579345703125,"n":8,"spread_pct":7.971349137518493},{"cp_plan":{"max_rank_tokens":10222,"summary_ms":0.0005364418029785156,"wave_count":0},"fallback_work":506296064,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1077.4367065429688,"n":8,"spread_pct":72.86943776322781},{"cp_plan":{"max_rank_tokens":53216,"summary_ms":0.0007841736078262329,"wave_count":0},"fallback_work":2624581632,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":2434.453369140625,"n":8,"spread_pct":0.8855118491953233},{"cp_plan":{"max_rank_tokens":9199,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":456112000,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1059.67236328125,"n":8,"spread_pct":10.966501383046518}],"cell":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":37871,"summary_ms":0.0008195638656616211,"wave_count":0},"fallback_work":1868609408,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":2077.63818359375,"n":8,"spread_pct":0.8851580128145621},{"cp_plan":{"max_rank_tokens":10222,"summary_ms":0.0003203749656677246,"wave_count":0},"fallback_work":506296064,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":914.9701538085938,"n":8,"spread_pct":6.308719486679593},{"cp_plan":{"max_rank_tokens":53216,"summary_ms":0.0015404075384140015,"wave_count":0},"fallback_work":2624581632,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":2677.4462890625,"n":8,"spread_pct":0.8127150805794205},{"cp_plan":{"max_rank_tokens":9199,"summary_ms":0.00028312206268310547,"wave_count":0},"fallback_work":456112000,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":905.1652221679688,"n":8,"spread_pct":4.222201049248943}],"cell":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":28852,"summary_ms":0.00040978193283081055,"wave_count":0},"fallback_work":1424841216,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":2470.6558837890625,"n":8,"spread_pct":0.8888905138650625},{"cp_plan":{"max_rank_tokens":33460,"summary_ms":0.00023283064365386963,"wave_count":0},"fallback_work":1651464704,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":2799.3975830078125,"n":8,"spread_pct":0.277778443979363},{"cp_plan":{"max_rank_tokens":29364,"summary_ms":0.00021420419216156006,"wave_count":0},"fallback_work":1449843200,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":2505.8414306640625,"n":8,"spread_pct":167.26044365185194},{"cp_plan":{"max_rank_tokens":36532,"summary_ms":0.00021792948246002197,"wave_count":0},"fallback_work":1802623488,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":3023.1268310546875,"n":8,"spread_pct":0.40736582469411525},{"cp_plan":{"max_rank_tokens":48820,"summary_ms":0.0009350478649139404,"wave_count":0},"fallback_work":2407946752,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":4072.2808837890625,"n":8,"spread_pct":18.384337588663115}],"cell":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":18432,"summary_ms":6.894262973219156,"wave_count":1},"fallback_work":1424242752,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":1775.5928344726562,"n":16,"spread_pct":46.87435730384531},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":4.902268992736936,"wave_count":1},"fallback_work":1650513984,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":2061.0057373046875,"n":16,"spread_pct":36.31729758758106},{"cp_plan":{"max_rank_tokens":18944,"summary_ms":6.2788561917841434,"wave_count":1},"fallback_work":1449154624,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":1807.5885620117188,"n":16,"spread_pct":39.825772497545294},{"cp_plan":{"max_rank_tokens":25088,"summary_ms":5.989300087094307,"wave_count":1},"fallback_work":1801459776,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":2258.7291259765625,"n":16,"spread_pct":1.2971917921081888},{"cp_plan":{"max_rank_tokens":30720,"summary_ms":7.917732931673527,"wave_count":1},"fallback_work":2406127680,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":2798.6976318359375,"n":16,"spread_pct":0.5809063585527431}],"cell":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":18432,"summary_ms":6.260422058403492,"wave_count":1},"fallback_work":1424242752,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":1709.4850463867188,"n":16,"spread_pct":38.2787856828944},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":4.452097928151488,"wave_count":1},"fallback_work":1650513984,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":1937.2064819335938,"n":16,"spread_pct":34.18290025098623},{"cp_plan":{"max_rank_tokens":18944,"summary_ms":5.752668017521501,"wave_count":1},"fallback_work":1449154624,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":1728.7608032226562,"n":16,"spread_pct":1.274381384533133},{"cp_plan":{"max_rank_tokens":25088,"summary_ms":5.445976974442601,"wave_count":1},"fallback_work":1801459776,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":2105.5108642578125,"n":16,"spread_pct":0.966028914658189},{"cp_plan":{"max_rank_tokens":30720,"summary_ms":7.114178966730833,"wave_count":1},"fallback_work":2406127680,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":2698.8433837890625,"n":16,"spread_pct":27.76845097624602}],"cell":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13312,"summary_ms":13.601833954453468,"wave_count":3},"fallback_work":1425123168,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":1486.32666015625,"n":16,"spread_pct":48.658211555399035},{"cp_plan":{"max_rank_tokens":16384,"summary_ms":4.904275061562657,"wave_count":2},"fallback_work":1651021664,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":1515.2017211914062,"n":16,"spread_pct":2.4278123743079827},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":13.438750989735126,"wave_count":3},"fallback_work":1449891680,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":1478.8178100585938,"n":16,"spread_pct":50.605476068691544},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":5.104751093313098,"wave_count":2},"fallback_work":1801762656,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":1623.783935546875,"n":16,"spread_pct":45.279369973232676},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":7.819026010110974,"wave_count":1},"fallback_work":2406004576,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":2005.5856323242188,"n":16,"spread_pct":36.35810814614884}],"cell":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13312,"summary_ms":12.31529749929905,"wave_count":3},"fallback_work":1425123168,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":1440.5418701171875,"n":8,"spread_pct":15.350745141119942},{"cp_plan":{"max_rank_tokens":16384,"summary_ms":5.2589233964681625,"wave_count":2},"fallback_work":1651021664,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":1498.6566772460938,"n":8,"spread_pct":59.39418972663141},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":12.275269255042076,"wave_count":3},"fallback_work":1449891680,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":1416.0592041015625,"n":8,"spread_pct":68.25557178973027},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":5.578400567173958,"wave_count":2},"fallback_work":1801762656,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":1534.9813232421875,"n":8,"spread_pct":6.607901488774921},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":6.361622363328934,"wave_count":1},"fallback_work":2406004576,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":1819.1531982421875,"n":8,"spread_pct":44.19536771775856}],"cell":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13312,"summary_ms":13.491021934896708,"wave_count":3},"fallback_work":1425123168,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":1435.8742065429688,"n":16,"spread_pct":80.70047633271376},{"cp_plan":{"max_rank_tokens":16384,"summary_ms":4.566923016682267,"wave_count":2},"fallback_work":1651021664,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":1437.3273315429688,"n":16,"spread_pct":1.89714533900153},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":13.300522929057479,"wave_count":3},"fallback_work":1449891680,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":1440.6353759765625,"n":16,"spread_pct":3.361189525055109},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":4.8843990080058575,"wave_count":2},"fallback_work":1801762656,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":1522.4942016601562,"n":16,"spread_pct":1.6880724823804163},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":7.672770880162716,"wave_count":1},"fallback_work":2406004576,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":1907.6583251953125,"n":16,"spread_pct":33.06689979361431}],"cell":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":28852,"summary_ms":0.00030174851417541504,"wave_count":0},"fallback_work":1424841216,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":1534.6395874023438,"n":8,"spread_pct":1.2424187768346813},{"cp_plan":{"max_rank_tokens":33460,"summary_ms":0.0002868473529815674,"wave_count":0},"fallback_work":1651464704,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":1693.6309204101562,"n":8,"spread_pct":1.6040175795050744},{"cp_plan":{"max_rank_tokens":29364,"summary_ms":0.00020116567611694336,"wave_count":0},"fallback_work":1449843200,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":1556.5745849609375,"n":8,"spread_pct":45.392970325983875},{"cp_plan":{"max_rank_tokens":36532,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":1802623488,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":1811.5222778320312,"n":8,"spread_pct":40.22413626208184},{"cp_plan":{"max_rank_tokens":48820,"summary_ms":0.0004637986421585083,"wave_count":0},"fallback_work":2407946752,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":2362.2747802734375,"n":8,"spread_pct":1.1703537999525109}],"cell":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":28852,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":1424841216,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":1632.4662475585938,"n":8,"spread_pct":3.2060402162883896},{"cp_plan":{"max_rank_tokens":33460,"summary_ms":0.00023283064365386963,"wave_count":0},"fallback_work":1651464704,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":1836.0621948242188,"n":8,"spread_pct":2.116498537017582},{"cp_plan":{"max_rank_tokens":29364,"summary_ms":0.00030547380447387695,"wave_count":0},"fallback_work":1449843200,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":1654.8679809570312,"n":8,"spread_pct":0.8741669281306627},{"cp_plan":{"max_rank_tokens":36532,"summary_ms":0.0003073364496231079,"wave_count":0},"fallback_work":1802623488,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":1966.8600463867188,"n":8,"spread_pct":35.82314185228969},{"cp_plan":{"max_rank_tokens":48820,"summary_ms":0.0004470348358154297,"wave_count":0},"fallback_work":2407946752,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":2579.250244140625,"n":8,"spread_pct":1.2977101187559332}],"cell":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16756,"summary_ms":0.00026263296604156494,"wave_count":0},"fallback_work":827636224,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1466.54296875,"n":8,"spread_pct":9.849183280727688},{"cp_plan":{"max_rank_tokens":17524,"summary_ms":0.00035390257835388184,"wave_count":0},"fallback_work":865253888,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1498.9694213867188,"n":8,"spread_pct":150.9961902128802},{"cp_plan":{"max_rank_tokens":25204,"summary_ms":0.0008549541234970093,"wave_count":0},"fallback_work":1243494912,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":2005.9140014648438,"n":8,"spread_pct":4.651772050389937}],"cell":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9588,"summary_ms":5.571075016632676,"wave_count":1},"fallback_work":827323456,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1277.3406372070312,"n":16,"spread_pct":57.40751757374791},{"cp_plan":{"max_rank_tokens":10356,"summary_ms":5.817441968247294,"wave_count":1},"fallback_work":864838720,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1302.2736206054688,"n":16,"spread_pct":63.841160837854424},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":9.026514831930399,"wave_count":1},"fallback_work":1242645568,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1465.2989501953125,"n":16,"spread_pct":3.088281983040437}],"cell":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9588,"summary_ms":6.039798958227038,"wave_count":1},"fallback_work":827323456,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1128.835693359375,"n":16,"spread_pct":65.10081404309223},{"cp_plan":{"max_rank_tokens":10356,"summary_ms":6.200105883181095,"wave_count":1},"fallback_work":864838720,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1146.96240234375,"n":16,"spread_pct":62.79251538647217},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":8.316880092024803,"wave_count":1},"fallback_work":1242645568,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1414.7180786132812,"n":16,"spread_pct":3.1517242459372574}],"cell":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":10.991174960508943,"wave_count":3},"fallback_work":827904352,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1274.7514038085938,"n":16,"spread_pct":61.79647753202493},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":11.220806976780295,"wave_count":3},"fallback_work":865270112,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1284.2142333984375,"n":16,"spread_pct":61.57113834295143},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.447112187743187,"wave_count":3},"fallback_work":1242761568,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1349.8079223632812,"n":16,"spread_pct":116.37185430345292}],"cell":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":9.96258482336998,"wave_count":3},"fallback_work":827904352,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1117.6480102539062,"n":8,"spread_pct":15.0649767673891},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":9.411407634615898,"wave_count":3},"fallback_work":865270112,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1115.4676513671875,"n":8,"spread_pct":279.2489552590302},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":4.414483904838562,"wave_count":3},"fallback_work":1242761568,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1188.9453125,"n":8,"spread_pct":13.912949288694682}],"cell":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":10.736441938206553,"wave_count":3},"fallback_work":827904352,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1075.541748046875,"n":16,"spread_pct":66.52321833632624},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":10.575474938377738,"wave_count":3},"fallback_work":865270112,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1090.6126098632812,"n":16,"spread_pct":123.51148777937071},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.387331988662481,"wave_count":3},"fallback_work":1242761568,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1217.6163940429688,"n":16,"spread_pct":5.476347766780469}],"cell":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16756,"summary_ms":0.0004246830940246582,"wave_count":0},"fallback_work":827636224,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1118.9140014648438,"n":8,"spread_pct":3.08805988036299},{"cp_plan":{"max_rank_tokens":17524,"summary_ms":0.00043585896492004395,"wave_count":0},"fallback_work":865253888,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1131.166748046875,"n":8,"spread_pct":63.12038675152846},{"cp_plan":{"max_rank_tokens":25204,"summary_ms":0.0003259629011154175,"wave_count":0},"fallback_work":1243494912,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1308.7330322265625,"n":8,"spread_pct":1.3643323970834822}],"cell":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16756,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":827636224,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1061.8983764648438,"n":8,"spread_pct":4.966851922070074},{"cp_plan":{"max_rank_tokens":17524,"summary_ms":0.0002998858690261841,"wave_count":0},"fallback_work":865253888,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1090.832275390625,"n":8,"spread_pct":64.96590570067606},{"cp_plan":{"max_rank_tokens":25204,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":1243494912,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1367.2909545898438,"n":8,"spread_pct":3.4822543561993857}],"cell":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19380,"summary_ms":0.0003781169652938843,"wave_count":0},"fallback_work":957307392,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1629.9077758789062,"n":8,"spread_pct":43.32433644197106},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":1234393600,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":2013.2527465820312,"n":8,"spread_pct":1.4115207599738948},{"cp_plan":{"max_rank_tokens":20916,"summary_ms":0.0003259629011154175,"wave_count":0},"fallback_work":1032772096,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1719.57373046875,"n":8,"spread_pct":0.7739472587995468},{"cp_plan":{"max_rank_tokens":29108,"summary_ms":0.0008400529623031616,"wave_count":0},"fallback_work":1436015104,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":2335.5113525390625,"n":8,"spread_pct":0.46186310431109684}],"cell":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11264,"summary_ms":7.4452729895710945,"wave_count":1},"fallback_work":956966976,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1377.8363037109375,"n":16,"spread_pct":60.52733020974579},{"cp_plan":{"max_rank_tokens":15872,"summary_ms":8.416526950895786,"wave_count":1},"fallback_work":1233651776,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1508.984619140625,"n":16,"spread_pct":4.083103468304513},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":9.881744859740138,"wave_count":1},"fallback_work":1032292416,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1412.7406616210938,"n":16,"spread_pct":55.03132827505408},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":10.069511830806732,"wave_count":1},"fallback_work":1435011136,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1626.6113891601562,"n":16,"spread_pct":4.4046541738715295}],"cell":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,2,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11264,"summary_ms":7.460470078513026,"wave_count":1},"fallback_work":956966976,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1344.5885009765625,"n":16,"spread_pct":61.950060413791576},{"cp_plan":{"max_rank_tokens":15872,"summary_ms":6.007199175655842,"wave_count":1},"fallback_work":1233651776,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1481.7916870117188,"n":16,"spread_pct":4.9539093014467},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":7.998563814908266,"wave_count":1},"fallback_work":1032292416,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1375.4036254882812,"n":16,"spread_pct":6.918423671176425},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":9.889952139928937,"wave_count":1},"fallback_work":1435011136,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1624.5980224609375,"n":16,"spread_pct":4.313597106014882}],"cell":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,2,2,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6580,"summary_ms":4.769534105435014,"wave_count":3},"fallback_work":957681504,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1338.1644287109375,"n":16,"spread_pct":160.08721937654548},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":4.943178966641426,"wave_count":2},"fallback_work":1233968992,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1352.83349609375,"n":16,"spread_pct":56.134955098222306},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.739928059279919,"wave_count":1},"fallback_work":1032839008,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1298.21728515625,"n":16,"spread_pct":115.02843066419584},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":12.874939944595098,"wave_count":3},"fallback_work":1435098976,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1340.1981201171875,"n":16,"spread_pct":3.0485012937995717}],"cell":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,4,1,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6580,"summary_ms":4.179641604423523,"wave_count":3},"fallback_work":957681504,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1271.4913330078125,"n":8,"spread_pct":88.04374822449608},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":4.411159083247185,"wave_count":2},"fallback_work":1233968992,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1271.0892944335938,"n":8,"spread_pct":3.395361855668767},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.228748381137848,"wave_count":1},"fallback_work":1032839008,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1220.8251953125,"n":8,"spread_pct":4.468643135686431},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":9.772999212145805,"wave_count":3},"fallback_work":1435098976,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1367.435791015625,"n":8,"spread_pct":59.33101970588556}],"cell":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,4,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6580,"summary_ms":4.718300886452198,"wave_count":3},"fallback_work":957681504,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1156.2494506835938,"n":16,"spread_pct":3.7091022688936035},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":4.922118037939072,"wave_count":2},"fallback_work":1233968992,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1232.530029296875,"n":16,"spread_pct":3.30618767013909},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.636699100956321,"wave_count":1},"fallback_work":1032839008,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1141.828857421875,"n":16,"spread_pct":62.098318320283106},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":12.140921084210277,"wave_count":3},"fallback_work":1435098976,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1325.11376953125,"n":16,"spread_pct":51.498213044059824}],"cell":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,4,4,1],"source":"5f9a99ce85680fd0f5fdf91e4b33e33e4067b98a9a5bd182f5617371bbfe45f8","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19380,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":957307392,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1190.34228515625,"n":8,"spread_pct":6.654210871101693},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.0004023313522338867,"wave_count":0},"fallback_work":1234393600,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1331.8233032226562,"n":8,"spread_pct":71.49350157440726},{"cp_plan":{"max_rank_tokens":20916,"summary_ms":0.0005029141902923584,"wave_count":0},"fallback_work":1032772096,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1224.0546875,"n":8,"spread_pct":73.68117097696565},{"cp_plan":{"max_rank_tokens":29108,"summary_ms":0.00042282044887542725,"wave_count":0},"fallback_work":1436015104,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1468.8363037109375,"n":8,"spread_pct":2.0180413268601987}],"cell":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19380,"summary_ms":0.00026263296604156494,"wave_count":0},"fallback_work":957307392,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1155.0253295898438,"n":8,"spread_pct":71.97463952396794},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.00023655593395233154,"wave_count":0},"fallback_work":1234393600,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1371.664306640625,"n":8,"spread_pct":1.9359198282338705},{"cp_plan":{"max_rank_tokens":20916,"summary_ms":0.0002998858690261841,"wave_count":0},"fallback_work":1032772096,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1197.3412475585938,"n":8,"spread_pct":2.4515376501249744},{"cp_plan":{"max_rank_tokens":29108,"summary_ms":0.0005476176738739014,"wave_count":0},"fallback_work":1436015104,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1557.0986938476562,"n":8,"spread_pct":2.5168409311798734}],"cell":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":0.0,"layers":48.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":128,"moe_ffn_hidden_size":768,"moe_shared_expert_ffn":0,"moe_topk":8,"num_attention_heads":32,"num_query_groups":4},"hidden_size":2048,"model":"Qwen/Qwen3-30B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[2,1,2,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero","seed":7777}} + ], + "fit_arguments": { + "exclude_cells": ["cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g6","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g7","cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp1"], + "holdout": "", + "incumbent": "depth_one", + "objective": "regret", + "require_complete": 8, + "rerank_shapes": ["tp1cp4","tp1cp4ep2","tp1cp4ep4"], + "shortlist_size": 3, + "terms": ["token_per_rank","token_cp_exchange","token_tp_collective","gdn_token_per_rank","attention_token_cp_exchange","tiny_segment_per_layer","level_cp_per_layer","level_tp_per_layer","gdn_level","gdn_level_tp"] + }, + "integer_table_milli_us": {"attention_token_cp_exchange":0,"gdn_level":0,"gdn_level_tp":0,"gdn_token_per_rank":0,"level_cp_per_layer":127543,"level_tp_per_layer":0,"tiny_segment_per_layer":0,"token_cp_exchange":237,"token_per_rank":6648,"token_tp_collective":105}, + "integer_table_sha256": "3ad3897ea39d9c8d73e3a52d1820eb12e876da14d338312601600d9c56274df4", + "manifest": { + "excluded": ["cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g6","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g7","cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp1"], + "expected": ["cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g0","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g1","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g2","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g3","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g4","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g5","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g6","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g7","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g0","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g1","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g2","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g3","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g4","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g5","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g6","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g7","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g7","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g0","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g1","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g2","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g3","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g4","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g5","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g6","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g7","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g0","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g1","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g2","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g3","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g4","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g5","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g6","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g7","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g0","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g1","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g2","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g3","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g4","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g5","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g6","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g7","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g0","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g1","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g2","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g3","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g4","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g5","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g6","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g7","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g0","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g1","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g2","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g3","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g4","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g5","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g6","cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g7","cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2"], + "path": "trainer_rank_cost_calibration_manifest_attn-moe-h2048-h200-bf16.json" + }, + "measured_envelope": {"device_names":["NVIDIA H200"],"hidden_sizes":[2048],"param_dtypes":["torch.bfloat16"]}, + "metrics": { + "integer_all": {"cells":67,"clear_misses":[],"max_regret_pct":2.154643306828303,"median_regret_pct":0.0,"ordered_pairs":3481,"p95_regret_pct":0.9705842165260201,"pairwise_accuracy":0.9629416834243033}, + "test": {"cells":19,"clear_misses":[],"max_regret_pct":2.0669357476129284,"median_regret_pct":0.03366143440493683,"ordered_pairs":854,"p95_regret_pct":2.0669357476129284,"pairwise_accuracy":0.9964871194379391}, + "train": {"cells":48,"clear_misses":[],"max_regret_pct":2.154643306828303,"median_regret_pct":0.0,"ordered_pairs":2627,"p95_regret_pct":0.6029008791204539,"pairwise_accuracy":0.9520365435858393} + }, + "reranker": {"incumbent":"depth_one","max_rank_token_per_layer_milli_us":1347,"metrics":{"all":{"cells":42,"clear_misses":[],"max_regret_pct":2.1008752938016557,"median_regret_pct":0.0,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":1.3827087677802716,"savings_vs_fallback_mean_pct":0.9127372971303251,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":80,"p95_regret_pct":1.639882683538256,"pairwise_accuracy":0.975,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g0":{"best":"depth_one","best_lead_pct":1.3038863967344374,"cheap_regret_pct":0.0,"fallback_regret_pct":2.978788739218399,"irrecoverable_pct":0.0,"planning_ms":8.605284616351128,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","uniform_depth_2","minimum_effective_span_46"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g1":{"best":"depth_one","best_lead_pct":11.961590492982479,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.71739536523819,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g2":{"best":"no_sharing","best_lead_pct":0.0,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":0.0,"regret_pct":0.0,"selected":"no_sharing","shortlist":["no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g3":{"best":"depth_one","best_lead_pct":0.86974850256131,"cheap_regret_pct":1.040523321231979,"fallback_regret_pct":1.3393167940631248,"irrecoverable_pct":0.0,"planning_ms":22.703632712364197,"regret_pct":1.040523321231979,"selected":"minimum_effective_span_278","shortlist":["minimum_effective_span_278","uniform_depth_3","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g4":{"best":"uniform_depth_2","best_lead_pct":5.334436917544552,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":13.415971770882607,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_6717"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g5":{"best":"minimum_effective_span_179","best_lead_pct":0.5561050302996507,"cheap_regret_pct":19.0872105543058,"fallback_regret_pct":1.3905311918791174,"irrecoverable_pct":0.0,"planning_ms":10.265113785862923,"regret_pct":0.5561050302996507,"selected":"minimum_effective_span_22","shortlist":["depth_one","minimum_effective_span_22","minimum_effective_span_179"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g6":{"best":"minimum_effective_span_71","best_lead_pct":0.3570250689122993,"cheap_regret_pct":0.4706833103497925,"fallback_regret_pct":10.404773566655116,"irrecoverable_pct":0.4706833103497925,"planning_ms":11.011766269803047,"regret_pct":0.4706833103497925,"selected":"minimum_effective_span_197","shortlist":["minimum_effective_span_197","minimum_effective_span_133","minimum_effective_span_164","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g7":{"best":"minimum_effective_span_134","best_lead_pct":1.025199621855049,"cheap_regret_pct":2.1008752938016557,"fallback_regret_pct":1.6682484028756863,"irrecoverable_pct":1.574728584874616,"planning_ms":9.171070531010628,"regret_pct":2.1008752938016557,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_87","minimum_effective_span_53"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g0":{"best":"depth_one","best_lead_pct":0.21704816254779188,"cheap_regret_pct":0.0,"fallback_regret_pct":1.530156236564887,"irrecoverable_pct":0.0,"planning_ms":9.267912246286869,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","uniform_depth_2","minimum_effective_span_46"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g1":{"best":"depth_one","best_lead_pct":16.117201547662415,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":10.899567976593971,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g2":{"best":"no_sharing","best_lead_pct":0.0,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":0.0,"regret_pct":0.0,"selected":"no_sharing","shortlist":["no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g3":{"best":"depth_one","best_lead_pct":0.3919510961307464,"cheap_regret_pct":1.5544362981525874,"fallback_regret_pct":0.3919510961307464,"irrecoverable_pct":0.0,"planning_ms":24.78388394229114,"regret_pct":1.5544362981525874,"selected":"minimum_effective_span_278","shortlist":["minimum_effective_span_278","uniform_depth_3","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g4":{"best":"uniform_depth_2","best_lead_pct":9.132792937693052,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":14.18487005867064,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_6717"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g5":{"best":"minimum_effective_span_179","best_lead_pct":0.7388976481493648,"cheap_regret_pct":18.150252978631073,"fallback_regret_pct":1.244211944225182,"irrecoverable_pct":0.0,"planning_ms":11.391336098313332,"regret_pct":1.137873652085147,"selected":"minimum_effective_span_22","shortlist":["depth_one","minimum_effective_span_22","minimum_effective_span_179"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g6":{"best":"uniform_depth_10","best_lead_pct":0.16054916029231514,"cheap_regret_pct":1.2850754898519239,"fallback_regret_pct":10.018343004123526,"irrecoverable_pct":1.2850754898519239,"planning_ms":11.508825700730085,"regret_pct":1.2850754898519239,"selected":"minimum_effective_span_197","shortlist":["minimum_effective_span_197","minimum_effective_span_133","minimum_effective_span_164","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g7":{"best":"minimum_effective_span_134","best_lead_pct":0.029767455802613876,"cheap_regret_pct":1.4792447975646474,"fallback_regret_pct":0.2944636279159617,"irrecoverable_pct":1.0970504406416115,"planning_ms":9.535392979159951,"regret_pct":1.4792447975646474,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_87","minimum_effective_span_53"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g0":{"best":"depth_one","best_lead_pct":0.4672059549913243,"cheap_regret_pct":0.0,"fallback_regret_pct":1.5885309265360272,"irrecoverable_pct":0.0,"planning_ms":8.943250169977546,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","uniform_depth_2","minimum_effective_span_46"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g1":{"best":"depth_one","best_lead_pct":7.022529724020789,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":11.491035111248493,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g2":{"best":"no_sharing","best_lead_pct":0.0,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":0.0,"regret_pct":0.0,"selected":"no_sharing","shortlist":["no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g3":{"best":"uniform_depth_3","best_lead_pct":0.5071595777985206,"cheap_regret_pct":1.639882683538256,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":25.49735293723643,"regret_pct":1.639882683538256,"selected":"minimum_effective_span_278","shortlist":["minimum_effective_span_278","uniform_depth_3","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g4":{"best":"uniform_depth_2","best_lead_pct":6.1378487199337926,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":14.770179986953735,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_6717"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g5":{"best":"minimum_effective_span_179","best_lead_pct":0.43644763619188576,"cheap_regret_pct":15.088220200826893,"fallback_regret_pct":0.43644763619188576,"irrecoverable_pct":0.0,"planning_ms":11.960187926888466,"regret_pct":1.04810687868614,"selected":"minimum_effective_span_22","shortlist":["depth_one","minimum_effective_span_22","minimum_effective_span_179"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g6":{"best":"minimum_effective_span_71","best_lead_pct":0.42672734833282533,"cheap_regret_pct":1.242379718554529,"fallback_regret_pct":10.384455944652128,"irrecoverable_pct":1.242379718554529,"planning_ms":12.011287966743112,"regret_pct":1.242379718554529,"selected":"minimum_effective_span_197","shortlist":["minimum_effective_span_197","minimum_effective_span_133","minimum_effective_span_164","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g7":{"best":"minimum_effective_span_134","best_lead_pct":0.4117308365345777,"cheap_regret_pct":1.0855234833275738,"fallback_regret_pct":0.815287251807164,"irrecoverable_pct":1.0855234833275738,"planning_ms":9.726435178890824,"regret_pct":1.0855234833275738,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_87","minimum_effective_span_53"]},"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":9.203397673936127,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":13.19465204142034,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2":{"best":"uniform_depth_2","best_lead_pct":10.129238896877949,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":16.804607585072517,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4":{"best":"uniform_depth_2","best_lead_pct":9.460789529833226,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":13.231751741841435,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":4.081130116688907,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":16.8946860358119,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2":{"best":"uniform_depth_2","best_lead_pct":2.128291280437947,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":13.98574560880661,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4":{"best":"uniform_depth_2","best_lead_pct":4.659134489284307,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":13.270851224660873,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":9.788418753247322,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":10.021115187555552,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2":{"best":"uniform_depth_2","best_lead_pct":10.697644178633528,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":9.543651714920998,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4":{"best":"uniform_depth_2","best_lead_pct":14.796977808723652,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":9.642817080020905,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4":{"best":"minimum_effective_span_257","best_lead_pct":0.5077603235897419,"cheap_regret_pct":0.5077603235897419,"fallback_regret_pct":0.5077603235897419,"irrecoverable_pct":0.0,"planning_ms":27.040584944188595,"regret_pct":0.5077603235897419,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_257","minimum_effective_span_1025"]},"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2":{"best":"minimum_effective_span_257","best_lead_pct":1.7289295493233527,"cheap_regret_pct":1.7289295493233527,"fallback_regret_pct":1.7289295493233527,"irrecoverable_pct":0.0,"planning_ms":24.590566754341125,"regret_pct":1.7289295493233527,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_257","minimum_effective_span_1025"]},"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4":{"best":"depth_one","best_lead_pct":0.10120141398030713,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":26.791544863954186,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_257","minimum_effective_span_1025"]},"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4":{"best":"depth_one","best_lead_pct":0.7423274500088027,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":22.211981937289238,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_769","no_sharing"]},"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2":{"best":"minimum_effective_span_769","best_lead_pct":0.1954659002478794,"cheap_regret_pct":0.1954659002478794,"fallback_regret_pct":0.1954659002478794,"irrecoverable_pct":0.0,"planning_ms":19.373992457985878,"regret_pct":0.1954659002478794,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_769","no_sharing"]},"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4":{"best":"depth_one","best_lead_pct":1.401234479626115,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":21.31191687658429,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_769","no_sharing"]},"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4":{"best":"minimum_effective_span_513","best_lead_pct":3.0770768508046453,"cheap_regret_pct":3.0770768508046453,"fallback_regret_pct":3.0770768508046453,"irrecoverable_pct":0.0,"planning_ms":9.71271307207644,"regret_pct":0.0,"selected":"minimum_effective_span_513","shortlist":["depth_one","minimum_effective_span_513","minimum_effective_span_2049"]},"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2":{"best":"minimum_effective_span_513","best_lead_pct":4.117223277672233,"cheap_regret_pct":4.15015498450155,"fallback_regret_pct":4.15015498450155,"irrecoverable_pct":0.0,"planning_ms":8.639907464385033,"regret_pct":0.0,"selected":"minimum_effective_span_513","shortlist":["depth_one","minimum_effective_span_513","minimum_effective_span_2049"]},"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4":{"best":"minimum_effective_span_513","best_lead_pct":1.2629382387723915,"cheap_regret_pct":1.2629382387723915,"fallback_regret_pct":1.2629382387723915,"irrecoverable_pct":0.0,"planning_ms":9.64041892439127,"regret_pct":0.0,"selected":"minimum_effective_span_513","shortlist":["depth_one","minimum_effective_span_513","minimum_effective_span_2049"]}},"planning_cost":{"max_cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g3","max_ms":24.78388394229114,"max_pct_of_cell":2.023949980979626,"mean_pct_of_cell":1.0633349780951258,"median_pct_of_cell":0.9738488070598215},"recall":{"best_in_shortlist":36,"cells":42,"clear_winner_cells":4,"clear_winners_in_shortlist":4,"max_irrecoverable_pct":1.574728584874616}},"by_shape":{"tp1cp4":{"cells":14,"clear_misses":[],"max_regret_pct":1.639882683538256,"median_regret_pct":0.0,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":1.2226564409246714,"savings_vs_fallback_mean_pct":0.8061361318489536,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":26,"p95_regret_pct":1.639882683538256,"pairwise_accuracy":0.9615384615384616,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g0":{"best":"depth_one","best_lead_pct":0.4672059549913243,"cheap_regret_pct":0.0,"fallback_regret_pct":1.5885309265360272,"irrecoverable_pct":0.0,"planning_ms":8.943250169977546,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","uniform_depth_2","minimum_effective_span_46"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g1":{"best":"depth_one","best_lead_pct":7.022529724020789,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":11.491035111248493,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g2":{"best":"no_sharing","best_lead_pct":0.0,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":0.0,"regret_pct":0.0,"selected":"no_sharing","shortlist":["no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g3":{"best":"uniform_depth_3","best_lead_pct":0.5071595777985206,"cheap_regret_pct":1.639882683538256,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":25.49735293723643,"regret_pct":1.639882683538256,"selected":"minimum_effective_span_278","shortlist":["minimum_effective_span_278","uniform_depth_3","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g4":{"best":"uniform_depth_2","best_lead_pct":6.1378487199337926,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":14.770179986953735,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_6717"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g5":{"best":"minimum_effective_span_179","best_lead_pct":0.43644763619188576,"cheap_regret_pct":15.088220200826893,"fallback_regret_pct":0.43644763619188576,"irrecoverable_pct":0.0,"planning_ms":11.960187926888466,"regret_pct":1.04810687868614,"selected":"minimum_effective_span_22","shortlist":["depth_one","minimum_effective_span_22","minimum_effective_span_179"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g6":{"best":"minimum_effective_span_71","best_lead_pct":0.42672734833282533,"cheap_regret_pct":1.242379718554529,"fallback_regret_pct":10.384455944652128,"irrecoverable_pct":1.242379718554529,"planning_ms":12.011287966743112,"regret_pct":1.242379718554529,"selected":"minimum_effective_span_197","shortlist":["minimum_effective_span_197","minimum_effective_span_133","minimum_effective_span_164","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g7":{"best":"minimum_effective_span_134","best_lead_pct":0.4117308365345777,"cheap_regret_pct":1.0855234833275738,"fallback_regret_pct":0.815287251807164,"irrecoverable_pct":1.0855234833275738,"planning_ms":9.726435178890824,"regret_pct":1.0855234833275738,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_87","minimum_effective_span_53"]},"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":9.203397673936127,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":13.19465204142034,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":4.081130116688907,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":16.8946860358119,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":9.788418753247322,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":10.021115187555552,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4":{"best":"minimum_effective_span_257","best_lead_pct":0.5077603235897419,"cheap_regret_pct":0.5077603235897419,"fallback_regret_pct":0.5077603235897419,"irrecoverable_pct":0.0,"planning_ms":27.040584944188595,"regret_pct":0.5077603235897419,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_257","minimum_effective_span_1025"]},"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4":{"best":"depth_one","best_lead_pct":0.7423274500088027,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":22.211981937289238,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_769","no_sharing"]},"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4":{"best":"minimum_effective_span_513","best_lead_pct":3.0770768508046453,"cheap_regret_pct":3.0770768508046453,"fallback_regret_pct":3.0770768508046453,"irrecoverable_pct":0.0,"planning_ms":9.71271307207644,"regret_pct":0.0,"selected":"minimum_effective_span_513","shortlist":["depth_one","minimum_effective_span_513","minimum_effective_span_2049"]}},"planning_cost":{"max_cell":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","max_ms":27.040584944188595,"max_pct_of_cell":1.8285271356798978,"mean_pct_of_cell":1.0231871201884795,"median_pct_of_cell":0.9738488070598215},"recall":{"best_in_shortlist":12,"cells":14,"clear_winner_cells":1,"clear_winners_in_shortlist":1,"max_irrecoverable_pct":1.242379718554529}},"tp1cp4ep2":{"cells":14,"clear_misses":[],"max_regret_pct":2.1008752938016557,"median_regret_pct":0.0,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":1.6200900363219783,"savings_vs_fallback_mean_pct":1.2688304802507082,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":28,"p95_regret_pct":2.1008752938016557,"pairwise_accuracy":1.0,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g0":{"best":"depth_one","best_lead_pct":1.3038863967344374,"cheap_regret_pct":0.0,"fallback_regret_pct":2.978788739218399,"irrecoverable_pct":0.0,"planning_ms":8.605284616351128,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","uniform_depth_2","minimum_effective_span_46"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g1":{"best":"depth_one","best_lead_pct":11.961590492982479,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.71739536523819,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g2":{"best":"no_sharing","best_lead_pct":0.0,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":0.0,"regret_pct":0.0,"selected":"no_sharing","shortlist":["no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g3":{"best":"depth_one","best_lead_pct":0.86974850256131,"cheap_regret_pct":1.040523321231979,"fallback_regret_pct":1.3393167940631248,"irrecoverable_pct":0.0,"planning_ms":22.703632712364197,"regret_pct":1.040523321231979,"selected":"minimum_effective_span_278","shortlist":["minimum_effective_span_278","uniform_depth_3","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g4":{"best":"uniform_depth_2","best_lead_pct":5.334436917544552,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":13.415971770882607,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_6717"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g5":{"best":"minimum_effective_span_179","best_lead_pct":0.5561050302996507,"cheap_regret_pct":19.0872105543058,"fallback_regret_pct":1.3905311918791174,"irrecoverable_pct":0.0,"planning_ms":10.265113785862923,"regret_pct":0.5561050302996507,"selected":"minimum_effective_span_22","shortlist":["depth_one","minimum_effective_span_22","minimum_effective_span_179"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g6":{"best":"minimum_effective_span_71","best_lead_pct":0.3570250689122993,"cheap_regret_pct":0.4706833103497925,"fallback_regret_pct":10.404773566655116,"irrecoverable_pct":0.4706833103497925,"planning_ms":11.011766269803047,"regret_pct":0.4706833103497925,"selected":"minimum_effective_span_197","shortlist":["minimum_effective_span_197","minimum_effective_span_133","minimum_effective_span_164","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g7":{"best":"minimum_effective_span_134","best_lead_pct":1.025199621855049,"cheap_regret_pct":2.1008752938016557,"fallback_regret_pct":1.6682484028756863,"irrecoverable_pct":1.574728584874616,"planning_ms":9.171070531010628,"regret_pct":2.1008752938016557,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_87","minimum_effective_span_53"]},"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2":{"best":"uniform_depth_2","best_lead_pct":10.129238896877949,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":16.804607585072517,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2":{"best":"uniform_depth_2","best_lead_pct":2.128291280437947,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":13.98574560880661,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2":{"best":"uniform_depth_2","best_lead_pct":10.697644178633528,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":9.543651714920998,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2":{"best":"minimum_effective_span_257","best_lead_pct":1.7289295493233527,"cheap_regret_pct":1.7289295493233527,"fallback_regret_pct":1.7289295493233527,"irrecoverable_pct":0.0,"planning_ms":24.590566754341125,"regret_pct":1.7289295493233527,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_257","minimum_effective_span_1025"]},"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2":{"best":"minimum_effective_span_769","best_lead_pct":0.1954659002478794,"cheap_regret_pct":0.1954659002478794,"fallback_regret_pct":0.1954659002478794,"irrecoverable_pct":0.0,"planning_ms":19.373992457985878,"regret_pct":0.1954659002478794,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_769","no_sharing"]},"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2":{"best":"minimum_effective_span_513","best_lead_pct":4.117223277672233,"cheap_regret_pct":4.15015498450155,"fallback_regret_pct":4.15015498450155,"irrecoverable_pct":0.0,"planning_ms":8.639907464385033,"regret_pct":0.0,"selected":"minimum_effective_span_513","shortlist":["depth_one","minimum_effective_span_513","minimum_effective_span_2049"]}},"planning_cost":{"max_cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g3","max_ms":22.703632712364197,"max_pct_of_cell":1.7871986359893468,"mean_pct_of_cell":1.0373638100607194,"median_pct_of_cell":1.0243195852666644},"recall":{"best_in_shortlist":12,"cells":14,"clear_winner_cells":2,"clear_winners_in_shortlist":2,"max_irrecoverable_pct":1.574728584874616}},"tp1cp4ep4":{"cells":14,"clear_misses":[],"max_regret_pct":1.5544362981525874,"median_regret_pct":0.0,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":1.3053798260941656,"savings_vs_fallback_mean_pct":0.6632452792913135,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":26,"p95_regret_pct":1.5544362981525874,"pairwise_accuracy":0.9615384615384616,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g0":{"best":"depth_one","best_lead_pct":0.21704816254779188,"cheap_regret_pct":0.0,"fallback_regret_pct":1.530156236564887,"irrecoverable_pct":0.0,"planning_ms":9.267912246286869,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","uniform_depth_2","minimum_effective_span_46"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g1":{"best":"depth_one","best_lead_pct":16.117201547662415,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":10.899567976593971,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g2":{"best":"no_sharing","best_lead_pct":0.0,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":0.0,"regret_pct":0.0,"selected":"no_sharing","shortlist":["no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g3":{"best":"depth_one","best_lead_pct":0.3919510961307464,"cheap_regret_pct":1.5544362981525874,"fallback_regret_pct":0.3919510961307464,"irrecoverable_pct":0.0,"planning_ms":24.78388394229114,"regret_pct":1.5544362981525874,"selected":"minimum_effective_span_278","shortlist":["minimum_effective_span_278","uniform_depth_3","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g4":{"best":"uniform_depth_2","best_lead_pct":9.132792937693052,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":14.18487005867064,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_6717"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g5":{"best":"minimum_effective_span_179","best_lead_pct":0.7388976481493648,"cheap_regret_pct":18.150252978631073,"fallback_regret_pct":1.244211944225182,"irrecoverable_pct":0.0,"planning_ms":11.391336098313332,"regret_pct":1.137873652085147,"selected":"minimum_effective_span_22","shortlist":["depth_one","minimum_effective_span_22","minimum_effective_span_179"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g6":{"best":"uniform_depth_10","best_lead_pct":0.16054916029231514,"cheap_regret_pct":1.2850754898519239,"fallback_regret_pct":10.018343004123526,"irrecoverable_pct":1.2850754898519239,"planning_ms":11.508825700730085,"regret_pct":1.2850754898519239,"selected":"minimum_effective_span_197","shortlist":["minimum_effective_span_197","minimum_effective_span_133","minimum_effective_span_164","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g7":{"best":"minimum_effective_span_134","best_lead_pct":0.029767455802613876,"cheap_regret_pct":1.4792447975646474,"fallback_regret_pct":0.2944636279159617,"irrecoverable_pct":1.0970504406416115,"planning_ms":9.535392979159951,"regret_pct":1.4792447975646474,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_87","minimum_effective_span_53"]},"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4":{"best":"uniform_depth_2","best_lead_pct":9.460789529833226,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":13.231751741841435,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4":{"best":"uniform_depth_2","best_lead_pct":4.659134489284307,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":13.270851224660873,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4":{"best":"uniform_depth_2","best_lead_pct":14.796977808723652,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":9.642817080020905,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4":{"best":"depth_one","best_lead_pct":0.10120141398030713,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":26.791544863954186,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_257","minimum_effective_span_1025"]},"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4":{"best":"depth_one","best_lead_pct":1.401234479626115,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":21.31191687658429,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_769","no_sharing"]},"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4":{"best":"minimum_effective_span_513","best_lead_pct":1.2629382387723915,"cheap_regret_pct":1.2629382387723915,"fallback_regret_pct":1.2629382387723915,"irrecoverable_pct":0.0,"planning_ms":9.64041892439127,"regret_pct":0.0,"selected":"minimum_effective_span_513","shortlist":["depth_one","minimum_effective_span_513","minimum_effective_span_2049"]}},"planning_cost":{"max_cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g3","max_ms":24.78388394229114,"max_pct_of_cell":2.023949980979626,"mean_pct_of_cell":1.1294540040361787,"median_pct_of_cell":1.1283987565197897},"recall":{"best_in_shortlist":12,"cells":14,"clear_winner_cells":1,"clear_winners_in_shortlist":1,"max_irrecoverable_pct":1.2850754898519239}}},"held_out":{"cells":12,"clear_misses":[],"max_regret_pct":2.1008752938016557,"median_regret_pct":1.066815181006857,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":4.131966514391069,"savings_vs_fallback_mean_pct":-0.3385094577998974,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":18,"p95_regret_pct":2.1008752938016557,"pairwise_accuracy":1.0,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g1":{"best":"depth_one","best_lead_pct":11.961590492982479,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.71739536523819,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g3":{"best":"depth_one","best_lead_pct":0.86974850256131,"cheap_regret_pct":1.040523321231979,"fallback_regret_pct":1.3393167940631248,"irrecoverable_pct":0.0,"planning_ms":22.703632712364197,"regret_pct":1.040523321231979,"selected":"minimum_effective_span_278","shortlist":["minimum_effective_span_278","uniform_depth_3","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g5":{"best":"minimum_effective_span_179","best_lead_pct":0.5561050302996507,"cheap_regret_pct":19.0872105543058,"fallback_regret_pct":1.3905311918791174,"irrecoverable_pct":0.0,"planning_ms":10.265113785862923,"regret_pct":0.5561050302996507,"selected":"minimum_effective_span_22","shortlist":["depth_one","minimum_effective_span_22","minimum_effective_span_179"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g7":{"best":"minimum_effective_span_134","best_lead_pct":1.025199621855049,"cheap_regret_pct":2.1008752938016557,"fallback_regret_pct":1.6682484028756863,"irrecoverable_pct":1.574728584874616,"planning_ms":9.171070531010628,"regret_pct":2.1008752938016557,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_87","minimum_effective_span_53"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g1":{"best":"depth_one","best_lead_pct":16.117201547662415,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":10.899567976593971,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g3":{"best":"depth_one","best_lead_pct":0.3919510961307464,"cheap_regret_pct":1.5544362981525874,"fallback_regret_pct":0.3919510961307464,"irrecoverable_pct":0.0,"planning_ms":24.78388394229114,"regret_pct":1.5544362981525874,"selected":"minimum_effective_span_278","shortlist":["minimum_effective_span_278","uniform_depth_3","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g5":{"best":"minimum_effective_span_179","best_lead_pct":0.7388976481493648,"cheap_regret_pct":18.150252978631073,"fallback_regret_pct":1.244211944225182,"irrecoverable_pct":0.0,"planning_ms":11.391336098313332,"regret_pct":1.137873652085147,"selected":"minimum_effective_span_22","shortlist":["depth_one","minimum_effective_span_22","minimum_effective_span_179"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g7":{"best":"minimum_effective_span_134","best_lead_pct":0.029767455802613876,"cheap_regret_pct":1.4792447975646474,"fallback_regret_pct":0.2944636279159617,"irrecoverable_pct":1.0970504406416115,"planning_ms":9.535392979159951,"regret_pct":1.4792447975646474,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_87","minimum_effective_span_53"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g1":{"best":"depth_one","best_lead_pct":7.022529724020789,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":11.491035111248493,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g3":{"best":"uniform_depth_3","best_lead_pct":0.5071595777985206,"cheap_regret_pct":1.639882683538256,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":25.49735293723643,"regret_pct":1.639882683538256,"selected":"minimum_effective_span_278","shortlist":["minimum_effective_span_278","uniform_depth_3","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g5":{"best":"minimum_effective_span_179","best_lead_pct":0.43644763619188576,"cheap_regret_pct":15.088220200826893,"fallback_regret_pct":0.43644763619188576,"irrecoverable_pct":0.0,"planning_ms":11.960187926888466,"regret_pct":1.04810687868614,"selected":"minimum_effective_span_22","shortlist":["depth_one","minimum_effective_span_22","minimum_effective_span_179"]},"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g7":{"best":"minimum_effective_span_134","best_lead_pct":0.4117308365345777,"cheap_regret_pct":1.0855234833275738,"fallback_regret_pct":0.815287251807164,"irrecoverable_pct":1.0855234833275738,"planning_ms":9.726435178890824,"regret_pct":1.0855234833275738,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_87","minimum_effective_span_53"]}},"planning_cost":{"max_cell":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g3","max_ms":24.78388394229114,"max_pct_of_cell":2.023949980979626,"mean_pct_of_cell":1.0794581849487084,"median_pct_of_cell":0.8572071078096248},"recall":{"best_in_shortlist":9,"cells":12,"clear_winner_cells":1,"clear_winners_in_shortlist":1,"max_irrecoverable_pct":1.574728584874616}}},"shapes":["tp1cp4","tp1cp4ep2","tp1cp4ep4"],"shortlist_size":3,"shortlist_table_milli_us":{"attention_token_cp_exchange":8,"gdn_level":0,"gdn_level_tp":0,"gdn_token_per_rank":0,"level_cp_per_layer":163376,"level_tp_per_layer":140748,"tiny_segment_per_layer":169817,"token_cp_exchange":98,"token_per_rank":6705,"token_tp_collective":78},"wave_per_layer_milli_us":1388921}, + "schema": "art.dev.trainer_rank_cost_calibration_certificate.v1", + "table_id": "attn-moe-h2048-h200-bf16" +} diff --git a/dev/trainer_rank_cost_calibration_certificate_dense-attn-h2048-h200-bf16.json b/dev/trainer_rank_cost_calibration_certificate_dense-attn-h2048-h200-bf16.json new file mode 100644 index 000000000..13c26d490 --- /dev/null +++ b/dev/trainer_rank_cost_calibration_certificate_dense-attn-h2048-h200-bf16.json @@ -0,0 +1,87 @@ +{ + "admitted": {"device_classes":[{"capability":[9,0],"memory_class":"hbm-141g"}],"geometries":[{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8}],"param_dtypes":["torch.bfloat16"],"shapes":[[1,1,1,1],[1,2,1,1],[1,4,1,1],[2,1,1,1]]}, + "cells": [ + {"candidates":[{"cp_plan":{"max_rank_tokens":12204,"summary_ms":0.0006612390279769897,"wave_count":0},"fallback_work":352228864,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":504.70616149902344,"n":8,"spread_pct":0.8253552689176819},{"cp_plan":{"max_rank_tokens":23926,"summary_ms":0.00038370490074157715,"wave_count":0},"fallback_work":689822464,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":940.9892883300781,"n":8,"spread_pct":1.8395466067585626},{"cp_plan":{"max_rank_tokens":35778,"summary_ms":0.00038370490074157715,"wave_count":0},"fallback_work":1031160064,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":1383.8194580078125,"n":8,"spread_pct":0.6887022329872959},{"cp_plan":{"max_rank_tokens":12154,"summary_ms":0.00037066638469696045,"wave_count":0},"fallback_work":351018240,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":508.6481628417969,"n":8,"spread_pct":1.6197203844178776},{"cp_plan":{"max_rank_tokens":47640,"summary_ms":0.0015962868928909302,"wave_count":0},"fallback_work":1372556288,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1851.1531982421875,"n":8,"spread_pct":0.5108130772842915},{"cp_plan":{"max_rank_tokens":12114,"summary_ms":0.0007059425115585327,"wave_count":0},"fallback_work":349866240,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":505.775390625,"n":8,"spread_pct":1.0397358441749016},{"cp_plan":{"max_rank_tokens":12109,"summary_ms":0.0005867332220077515,"wave_count":0},"fallback_work":349951616,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":507.8727264404297,"n":8,"spread_pct":1.3224393128615426}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11920,"summary_ms":0.0004246830940246582,"wave_count":0},"fallback_work":343787520,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":478.0967254638672,"n":8,"spread_pct":2.5405275862188925},{"cp_plan":{"max_rank_tokens":23728,"summary_ms":0.0009406358003616333,"wave_count":0},"fallback_work":683628544,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":915.4790649414062,"n":8,"spread_pct":1.019407090067216}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12138,"summary_ms":0.0007152557373046875,"wave_count":0},"fallback_work":349705472,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":489.6867980957031,"n":8,"spread_pct":1.3434889331379043}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":20701,"summary_ms":0.00035762786865234375,"wave_count":0},"fallback_work":597073536,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":846.0764770507812,"n":8,"spread_pct":1.168905153520202},{"cp_plan":{"max_rank_tokens":16154,"summary_ms":0.0003129243850708008,"wave_count":0},"fallback_work":466349312,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":660.6011657714844,"n":8,"spread_pct":1.4518968563115677},{"cp_plan":{"max_rank_tokens":14771,"summary_ms":0.00022351741790771484,"wave_count":0},"fallback_work":426748288,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":604.4955749511719,"n":8,"spread_pct":1.8638122534753567},{"cp_plan":{"max_rank_tokens":21963,"summary_ms":0.00023096799850463867,"wave_count":0},"fallback_work":633648512,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":842.2384948730469,"n":8,"spread_pct":1.0260133613923756},{"cp_plan":{"max_rank_tokens":27497,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":793027712,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":1009.5647277832031,"n":8,"spread_pct":0.8512090035321038},{"cp_plan":{"max_rank_tokens":35243,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":1015883136,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":1274.1929321289062,"n":8,"spread_pct":0.3851051358193043},{"cp_plan":{"max_rank_tokens":46153,"summary_ms":0.000996515154838562,"wave_count":0},"fallback_work":1329861760,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":1711.7232055664062,"n":8,"spread_pct":0.7887573603916836},{"cp_plan":{"max_rank_tokens":15600,"summary_ms":0.0003557652235031128,"wave_count":0},"fallback_work":450623488,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":632.306640625,"n":8,"spread_pct":1.01254745305661},{"cp_plan":{"max_rank_tokens":14494,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":419000064,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":598.0364379882812,"n":8,"spread_pct":2.0461861015903997}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16800,"summary_ms":0.0003725290298461914,"wave_count":0},"fallback_work":484462592,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":684.185546875,"n":8,"spread_pct":1.459352688957845},{"cp_plan":{"max_rank_tokens":18971,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":546987392,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":719.5334777832031,"n":8,"spread_pct":1.0373288322388727},{"cp_plan":{"max_rank_tokens":30232,"summary_ms":0.0008344650268554688,"wave_count":0},"fallback_work":871074816,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":1141.8434448242188,"n":8,"spread_pct":0.737204793488206},{"cp_plan":{"max_rank_tokens":12255,"summary_ms":0.0002980232238769531,"wave_count":0},"fallback_work":353795968,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":497.4389343261719,"n":8,"spread_pct":2.215161043084515}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13078,"summary_ms":0.0006444752216339111,"wave_count":0},"fallback_work":377531136,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":561.5518188476562,"n":8,"spread_pct":1.0225244050164857},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.00034831464290618896,"wave_count":0},"fallback_work":721230336,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":995.1815795898438,"n":8,"spread_pct":1.198560814422276},{"cp_plan":{"max_rank_tokens":36716,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":1058305536,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":1422.7989501953125,"n":8,"spread_pct":0.8705709691221425},{"cp_plan":{"max_rank_tokens":48890,"summary_ms":0.00028312206268310547,"wave_count":0},"fallback_work":1408916736,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":1883.1594848632812,"n":8,"spread_pct":0.9265149365464773},{"cp_plan":{"max_rank_tokens":12899,"summary_ms":0.0002942979335784912,"wave_count":0},"fallback_work":372605312,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":553.4660339355469,"n":8,"spread_pct":0.663587152055514},{"cp_plan":{"max_rank_tokens":12722,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":367507712,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":543.1344604492188,"n":8,"spread_pct":0.8352666649061128},{"cp_plan":{"max_rank_tokens":61066,"summary_ms":0.0011511147022247314,"wave_count":0},"fallback_work":1759356160,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2399.294189453125,"n":8,"spread_pct":0.5189616208949797},{"cp_plan":{"max_rank_tokens":13015,"summary_ms":0.0002998858690261841,"wave_count":0},"fallback_work":375946112,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":558.6249084472656,"n":8,"spread_pct":0.6559948770340116},{"cp_plan":{"max_rank_tokens":12701,"summary_ms":0.0003241002559661865,"wave_count":0},"fallback_work":367132288,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":543.9251098632812,"n":8,"spread_pct":156.450233957298},{"cp_plan":{"max_rank_tokens":12700,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":367332864,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":537.9688415527344,"n":8,"spread_pct":1.1234182484017194}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9058,"summary_ms":0.0006593763828277588,"wave_count":0},"fallback_work":262803712,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":385.4941101074219,"n":8,"spread_pct":5.182280598020673},{"cp_plan":{"max_rank_tokens":6403,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":187257216,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":314.15000915527344,"n":8,"spread_pct":2.8700996957713754},{"cp_plan":{"max_rank_tokens":6443,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":188409216,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":315.33448791503906,"n":8,"spread_pct":1.3246354682994737},{"cp_plan":{"max_rank_tokens":6668,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":194659840,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":319.29864501953125,"n":8,"spread_pct":4.297223930394011},{"cp_plan":{"max_rank_tokens":6755,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":197165440,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":326.32733154296875,"n":8,"spread_pct":2.9549258614952794},{"cp_plan":{"max_rank_tokens":6799,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":198432640,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":326.3471984863281,"n":8,"spread_pct":3.8415728072812656},{"cp_plan":{"max_rank_tokens":7040,"summary_ms":0.00028312206268310547,"wave_count":0},"fallback_work":205144064,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":329.7487030029297,"n":8,"spread_pct":3.830598131613508},{"cp_plan":{"max_rank_tokens":7291,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":212372864,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":334.9156951904297,"n":8,"spread_pct":3.3778614868489663},{"cp_plan":{"max_rank_tokens":7367,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":214561664,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":342.6578674316406,"n":8,"spread_pct":3.4685394728439864},{"cp_plan":{"max_rank_tokens":7666,"summary_ms":0.00029616057872772217,"wave_count":0},"fallback_work":222943488,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":344.8207092285156,"n":8,"spread_pct":2.5375417566997682},{"cp_plan":{"max_rank_tokens":5988,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":176681472,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":304.1422119140625,"n":8,"spread_pct":4.621793088833261},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.0003129243850708008,"wave_count":0},"fallback_work":225017088,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":349.2855682373047,"n":8,"spread_pct":3.936641911574962},{"cp_plan":{"max_rank_tokens":6024,"summary_ms":0.00027008354663848877,"wave_count":0},"fallback_work":177488896,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":302.6211242675781,"n":8,"spread_pct":3.5834877341715234},{"cp_plan":{"max_rank_tokens":6102,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":179276544,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":307.14637756347656,"n":8,"spread_pct":1.7060436496433407},{"cp_plan":{"max_rank_tokens":8016,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":233023488,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":351.30548095703125,"n":8,"spread_pct":2.1573845045396904},{"cp_plan":{"max_rank_tokens":13361,"summary_ms":0.00027567148208618164,"wave_count":0},"fallback_work":386730112,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":510.6410675048828,"n":8,"spread_pct":0.8502041797053929},{"cp_plan":{"max_rank_tokens":17592,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":508582912,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":629.7809143066406,"n":8,"spread_pct":1.0447717293025263},{"cp_plan":{"max_rank_tokens":22066,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":637434112,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":768.7078552246094,"n":8,"spread_pct":1.0583262492278878},{"cp_plan":{"max_rank_tokens":26682,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":770374912,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":914.5503540039062,"n":8,"spread_pct":0.5291646892787524},{"cp_plan":{"max_rank_tokens":31360,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":905101312,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":1042.6509399414062,"n":8,"spread_pct":0.6670337807172398},{"cp_plan":{"max_rank_tokens":35986,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":1038330112,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":1194.6433715820312,"n":8,"spread_pct":0.966145417290122},{"cp_plan":{"max_rank_tokens":40782,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":1176454912,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":1350.5742797851562,"n":8,"spread_pct":0.5012244999041704},{"cp_plan":{"max_rank_tokens":45522,"summary_ms":0.00027567148208618164,"wave_count":0},"fallback_work":1312966912,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":1499.6110229492188,"n":8,"spread_pct":0.9717852045535734},{"cp_plan":{"max_rank_tokens":50326,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":1451322112,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":1651.3985595703125,"n":8,"spread_pct":0.38276179998650234},{"cp_plan":{"max_rank_tokens":55336,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":1595610112,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":1772.1935424804688,"n":8,"spread_pct":0.37648930439540484},{"cp_plan":{"max_rank_tokens":8411,"summary_ms":0.00026263296604156494,"wave_count":0},"fallback_work":244399488,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":364.7945556640625,"n":8,"spread_pct":2.6812367968512882},{"cp_plan":{"max_rank_tokens":60463,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":1743267712,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":1953.8841552734375,"n":8,"spread_pct":0.4105091902091306},{"cp_plan":{"max_rank_tokens":6223,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":182531968,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":311.0763702392578,"n":8,"spread_pct":1.9411576203526015},{"cp_plan":{"max_rank_tokens":6304,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":184635392,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":308.78065490722656,"n":8,"spread_pct":1.186682373793681},{"cp_plan":{"max_rank_tokens":6322,"summary_ms":0.00032782554626464844,"wave_count":0},"fallback_work":185153792,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":312.3652038574219,"n":8,"spread_pct":3.018811923072677},{"cp_plan":{"max_rank_tokens":6333,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":185470592,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":312.74851989746094,"n":8,"spread_pct":1.6528944974277537},{"cp_plan":{"max_rank_tokens":6310,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":184808192,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":312.08091735839844,"n":8,"spread_pct":1.830022537548524},{"cp_plan":{"max_rank_tokens":65842,"summary_ms":0.0011827796697616577,"wave_count":0},"fallback_work":1897953536,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":2178.212646484375,"n":8,"spread_pct":0.5988146773595351},{"cp_plan":{"max_rank_tokens":6245,"summary_ms":0.0002905726432800293,"wave_count":0},"fallback_work":183853696,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":310.39581298828125,"n":8,"spread_pct":3.1717079357588447},{"cp_plan":{"max_rank_tokens":6083,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":179417472,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":307.4862365722656,"n":8,"spread_pct":4.107085665137808},{"cp_plan":{"max_rank_tokens":5957,"summary_ms":0.0002868473529815674,"wave_count":0},"fallback_work":176018048,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":304.5399475097656,"n":8,"spread_pct":4.421542283347354},{"cp_plan":{"max_rank_tokens":8629,"summary_ms":0.0003725290298461914,"wave_count":0},"fallback_work":250677888,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":372.5790710449219,"n":8,"spread_pct":3.0376688830422247},{"cp_plan":{"max_rank_tokens":8089,"summary_ms":0.00031478703022003174,"wave_count":0},"fallback_work":235355264,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":356.0412139892578,"n":8,"spread_pct":3.2818284530360344},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":225475840,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":346.13316345214844,"n":8,"spread_pct":1.6768335892899473},{"cp_plan":{"max_rank_tokens":7490,"summary_ms":0.0003259629011154175,"wave_count":0},"fallback_work":218562816,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":340.8607940673828,"n":8,"spread_pct":2.638422289830994},{"cp_plan":{"max_rank_tokens":7273,"summary_ms":0.00032782554626464844,"wave_count":0},"fallback_work":212542592,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":335.3833923339844,"n":8,"spread_pct":3.753602303508794},{"cp_plan":{"max_rank_tokens":6973,"summary_ms":0.0002942979335784912,"wave_count":0},"fallback_work":204131968,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":326.033935546875,"n":8,"spread_pct":2.8362668344022737},{"cp_plan":{"max_rank_tokens":6793,"summary_ms":0.0003259629011154175,"wave_count":0},"fallback_work":199177344,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":323.2712097167969,"n":8,"spread_pct":3.130685679561814},{"cp_plan":{"max_rank_tokens":6521,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":191573120,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":315.90943908691406,"n":8,"spread_pct":3.037707596529485}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12291,"summary_ms":0.0005066394805908203,"wave_count":0},"fallback_work":355258752,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":550.95556640625,"n":8,"spread_pct":1.206909692588427},{"cp_plan":{"max_rank_tokens":23563,"summary_ms":0.00027567148208618164,"wave_count":0},"fallback_work":679892352,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":950.5731506347656,"n":8,"spread_pct":1.2195625833650872},{"cp_plan":{"max_rank_tokens":34902,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":1006455552,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":1377.7364501953125,"n":8,"spread_pct":0.7508671706884479},{"cp_plan":{"max_rank_tokens":46161,"summary_ms":0.0005923211574554443,"wave_count":0},"fallback_work":1330714752,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":1780.816162109375,"n":8,"spread_pct":1.079635212688178},{"cp_plan":{"max_rank_tokens":57385,"summary_ms":0.0002905726432800293,"wave_count":0},"fallback_work":1653965952,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":2187.3680419921875,"n":8,"spread_pct":0.9042063307501976},{"cp_plan":{"max_rank_tokens":68715,"summary_ms":0.00028312206268310547,"wave_count":0},"fallback_work":1980269952,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":2621.7581787109375,"n":8,"spread_pct":0.8718263156372944},{"cp_plan":{"max_rank_tokens":80181,"summary_ms":0.0002905726432800293,"wave_count":0},"fallback_work":2310490752,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":3042.7430419921875,"n":8,"spread_pct":1.039390317422052},{"cp_plan":{"max_rank_tokens":11890,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":344627456,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":536.60498046875,"n":8,"spread_pct":0.8481942587903003},{"cp_plan":{"max_rank_tokens":12153,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":351513728,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":545.5311584472656,"n":8,"spread_pct":0.6494754269210081},{"cp_plan":{"max_rank_tokens":11903,"summary_ms":0.0003110617399215698,"wave_count":0},"fallback_work":344772480,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":536.7933959960938,"n":8,"spread_pct":1.1691423465827566},{"cp_plan":{"max_rank_tokens":11876,"summary_ms":0.00031851232051849365,"wave_count":0},"fallback_work":344453632,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":532.0890808105469,"n":8,"spread_pct":0.8563106467582366},{"cp_plan":{"max_rank_tokens":11939,"summary_ms":0.0003129243850708008,"wave_count":0},"fallback_work":345579904,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":537.6875305175781,"n":8,"spread_pct":19.775020581538794},{"cp_plan":{"max_rank_tokens":11981,"summary_ms":0.0002998858690261841,"wave_count":0},"fallback_work":346789504,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":538.9698791503906,"n":8,"spread_pct":0.7642164513660493},{"cp_plan":{"max_rank_tokens":12033,"summary_ms":0.00030547380447387695,"wave_count":0},"fallback_work":348057728,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":544.5603637695312,"n":8,"spread_pct":2.2709174108226144},{"cp_plan":{"max_rank_tokens":11881,"summary_ms":0.00031478703022003174,"wave_count":0},"fallback_work":344368256,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":536.1080017089844,"n":8,"spread_pct":216.24497991624955},{"cp_plan":{"max_rank_tokens":12025,"summary_ms":0.0002868473529815674,"wave_count":0},"fallback_work":347827328,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":541.4726257324219,"n":8,"spread_pct":142.5747035094171},{"cp_plan":{"max_rank_tokens":91657,"summary_ms":0.0010672956705093384,"wave_count":0},"fallback_work":2640770176,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3540.8409423828125,"n":8,"spread_pct":0.9949815160834775},{"cp_plan":{"max_rank_tokens":12225,"summary_ms":0.00036694109439849854,"wave_count":0},"fallback_work":353587328,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":548.0858764648438,"n":8,"spread_pct":0.7469733031961932},{"cp_plan":{"max_rank_tokens":12215,"summary_ms":0.0003129243850708008,"wave_count":0},"fallback_work":353528704,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":548.0600891113281,"n":8,"spread_pct":1.4207716058164803},{"cp_plan":{"max_rank_tokens":12123,"summary_ms":0.0002942979335784912,"wave_count":0},"fallback_work":351108480,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":544.4574279785156,"n":8,"spread_pct":1.1136507981349024},{"cp_plan":{"max_rank_tokens":11973,"summary_ms":0.00029616057872772217,"wave_count":0},"fallback_work":347017856,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":540.4607849121094,"n":8,"spread_pct":3.2455328877161307},{"cp_plan":{"max_rank_tokens":11879,"summary_ms":0.000303611159324646,"wave_count":0},"fallback_work":344540032,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":538.1383056640625,"n":8,"spread_pct":2.1105477710994096},{"cp_plan":{"max_rank_tokens":11874,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":344625408,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":537.2904357910156,"n":8,"spread_pct":0.5397839466956963}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.9595368802547455,"wave_count":2},"fallback_work":351839680,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":530.4389953613281,"n":8,"spread_pct":2.800990137459961},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":12.25663349032402,"wave_count":1},"fallback_work":688870624,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":665.1368103027344,"n":8,"spread_pct":0.5154253438245103},{"cp_plan":{"max_rank_tokens":20480,"summary_ms":5.201471969485283,"wave_count":1},"fallback_work":1029639328,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":883.7350769042969,"n":8,"spread_pct":1.7810618335077169},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.41260838508606,"wave_count":2},"fallback_work":350696992,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":523.4087524414062,"n":8,"spread_pct":3.1912550234510313},{"cp_plan":{"max_rank_tokens":27136,"summary_ms":8.616087958216667,"wave_count":1},"fallback_work":1370400640,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1148.0265502929688,"n":8,"spread_pct":0.6251280286467303},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.567955017089844,"wave_count":2},"fallback_work":349546912,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":527.4560241699219,"n":8,"spread_pct":2.9173071311958525},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.415927618741989,"wave_count":2},"fallback_work":349698144,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":526.4613342285156,"n":8,"spread_pct":1.676068490764494}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.322083994746208,"wave_count":2},"fallback_work":343346432,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":503.68040466308594,"n":8,"spread_pct":1.7987046580560768},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":8.494991809129715,"wave_count":1},"fallback_work":682555136,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":659.9745483398438,"n":8,"spread_pct":0.8211764647129669}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.239687860012054,"wave_count":2},"fallback_work":349155616,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":487.5807342529297,"n":8,"spread_pct":3.674966063030865}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13824,"summary_ms":8.759619668126106,"wave_count":1},"fallback_work":596309344,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":657.5940856933594,"n":8,"spread_pct":1.4251410209270514},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.337523460388184,"wave_count":2},"fallback_work":465868832,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":581.6876525878906,"n":8,"spread_pct":0.620364129600208},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":5.87029941380024,"wave_count":2},"fallback_work":426399808,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":564.1097106933594,"n":8,"spread_pct":5.015583929820624},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.184529513120651,"wave_count":1},"fallback_work":632889280,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":584.3733215332031,"n":8,"spread_pct":2.249484235128725},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":7.586374878883362,"wave_count":4},"fallback_work":792002848,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":678.4897155761719,"n":8,"spread_pct":2.0290426497377676},{"cp_plan":{"max_rank_tokens":17835,"summary_ms":5.464807152748108,"wave_count":1},"fallback_work":1014420928,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":786.2625732421875,"n":8,"spread_pct":1.7316706885183393},{"cp_plan":{"max_rank_tokens":24649,"summary_ms":6.47403672337532,"wave_count":1},"fallback_work":1327810336,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":1063.0380859375,"n":8,"spread_pct":103.58589078508953},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":7.0283859968185425,"wave_count":3},"fallback_work":450235136,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":575.8275146484375,"n":8,"spread_pct":4.0837966928963265},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":7.2373803704977036,"wave_count":3},"fallback_work":418730336,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":568.66455078125,"n":8,"spread_pct":4.807051626059891}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11776,"summary_ms":5.969293415546417,"wave_count":2},"fallback_work":483820032,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":583.2732238769531,"n":8,"spread_pct":2.5386474760731046},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":6.16094283759594,"wave_count":1},"fallback_work":546240704,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":596.6035461425781,"n":8,"spread_pct":0.967942666579026},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":7.224198430776596,"wave_count":1},"fallback_work":869721984,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":752.5882568359375,"n":8,"spread_pct":2.023274821607658},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.715016275644302,"wave_count":2},"fallback_work":353437184,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":519.9942016601562,"n":8,"spread_pct":1.3693272618325893}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.656464025378227,"wave_count":2},"fallback_work":377132768,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":552.6433410644531,"n":8,"spread_pct":2.1262114890057258},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":7.667737081646919,"wave_count":1},"fallback_work":720259136,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":659.7073059082031,"n":8,"spread_pct":1.4649940610036716},{"cp_plan":{"max_rank_tokens":21504,"summary_ms":5.994867533445358,"wave_count":1},"fallback_work":1056772544,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":949.2568054199219,"n":8,"spread_pct":1.0926066589344234},{"cp_plan":{"max_rank_tokens":25850,"summary_ms":8.103569969534874,"wave_count":1},"fallback_work":1406799392,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":1205.29736328125,"n":8,"spread_pct":1.4582638733877022},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.647666752338409,"wave_count":2},"fallback_work":372281152,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":546.5966186523438,"n":8,"spread_pct":1.7191340368685442},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.46296127140522,"wave_count":2},"fallback_work":367191968,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":547.1887512207031,"n":8,"spread_pct":3.028340715909583},{"cp_plan":{"max_rank_tokens":33280,"summary_ms":6.395192816853523,"wave_count":1},"fallback_work":1756588832,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1417.8174438476562,"n":8,"spread_pct":0.7156497114167787},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.625458434224129,"wave_count":2},"fallback_work":375616384,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":547.9529724121094,"n":8,"spread_pct":1.6288428785157472},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.529884248971939,"wave_count":2},"fallback_work":366883168,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":548.9385986328125,"n":8,"spread_pct":1.5133964359931997},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.574513226747513,"wave_count":2},"fallback_work":367149248,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":548.8893737792969,"n":8,"spread_pct":1.769197576760645}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7680,"summary_ms":10.019851848483086,"wave_count":1},"fallback_work":262860448,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":467.1503143310547,"n":8,"spread_pct":3.4032417204507457},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.108023062348366,"wave_count":2},"fallback_work":187703616,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":498.1083679199219,"n":8,"spread_pct":3.135280784878583},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.097804591059685,"wave_count":2},"fallback_work":188853696,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":496.6844024658203,"n":8,"spread_pct":3.471826066830065},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.223410367965698,"wave_count":2},"fallback_work":195027904,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":500.7582092285156,"n":8,"spread_pct":9.112786427076372},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.192248314619064,"wave_count":2},"fallback_work":197529408,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":500.50254821777344,"n":8,"spread_pct":3.8089462998539525},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.179794669151306,"wave_count":2},"fallback_work":198794496,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":506.95848083496094,"n":8,"spread_pct":7.34021106589986},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.777379661798477,"wave_count":2},"fallback_work":205428736,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":497.8880310058594,"n":8,"spread_pct":7.527547848440633},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.898213014006615,"wave_count":2},"fallback_work":212645568,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":504.9750518798828,"n":8,"spread_pct":3.8635712255784993},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.884796380996704,"wave_count":2},"fallback_work":214830720,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":502.13311767578125,"n":8,"spread_pct":3.955998754823933},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.173919886350632,"wave_count":2},"fallback_work":223132576,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":503.5539855957031,"n":8,"spread_pct":2.5887040583178997},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.644358694553375,"wave_count":1},"fallback_work":177540928,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":427.7280578613281,"n":8,"spread_pct":5.844956471603141},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":7.5253210961818695,"wave_count":1},"fallback_work":225202720,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":453.7100067138672,"n":8,"spread_pct":3.0592776923677367},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.683807656168938,"wave_count":1},"fallback_work":178281088,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":425.8500213623047,"n":8,"spread_pct":3.7543677990370465},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.178406834602356,"wave_count":1},"fallback_work":179933920,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":425.30078125,"n":8,"spread_pct":3.3555640768941095},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.908724084496498,"wave_count":1},"fallback_work":233195776,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":451.9730529785156,"n":8,"spread_pct":3.271382332226195},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":5.75120747089386,"wave_count":2},"fallback_work":386580384,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":544.5678100585938,"n":8,"spread_pct":2.220741770050186},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":6.347715854644775,"wave_count":1},"fallback_work":508230016,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":544.2649536132812,"n":8,"spread_pct":1.7973953142660668},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":7.225152105093002,"wave_count":1},"fallback_work":636866464,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":562.2056274414062,"n":8,"spread_pct":7.719990615748672},{"cp_plan":{"max_rank_tokens":16896,"summary_ms":7.569890469312668,"wave_count":1},"fallback_work":769585696,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":641.8765258789062,"n":8,"spread_pct":0.8674269636157641},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":8.130919188261032,"wave_count":1},"fallback_work":904087552,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":704.2494812011719,"n":8,"spread_pct":2.11263696465986},{"cp_plan":{"max_rank_tokens":20992,"summary_ms":7.889678701758385,"wave_count":1},"fallback_work":1037094304,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":782.9495239257812,"n":8,"spread_pct":0.8087952419131974},{"cp_plan":{"max_rank_tokens":23040,"summary_ms":6.783021613955498,"wave_count":1},"fallback_work":1174988896,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":858.9280395507812,"n":8,"spread_pct":1.2681098272915663},{"cp_plan":{"max_rank_tokens":26112,"summary_ms":9.522899985313416,"wave_count":2},"fallback_work":1311273376,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":968.908203125,"n":8,"spread_pct":1.4055605334612693},{"cp_plan":{"max_rank_tokens":27648,"summary_ms":8.692113682627678,"wave_count":1},"fallback_work":1449397984,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":1029.2799072265625,"n":8,"spread_pct":57.818102077346865},{"cp_plan":{"max_rank_tokens":30720,"summary_ms":11.38845644891262,"wave_count":1},"fallback_work":1593445504,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":1130.3540649414062,"n":8,"spread_pct":0.7466095196574506},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.316673010587692,"wave_count":1},"fallback_work":244552896,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":452.8104705810547,"n":8,"spread_pct":4.692959859293495},{"cp_plan":{"max_rank_tokens":32768,"summary_ms":9.331762790679932,"wave_count":1},"fallback_work":1740857088,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":1218.9605102539062,"n":8,"spread_pct":1.2323091829741548},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.13818334043026,"wave_count":1},"fallback_work":183118080,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":433.7216339111328,"n":8,"spread_pct":151.41326032746326},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.36767952144146,"wave_count":1},"fallback_work":185152000,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":435.29954528808594,"n":8,"spread_pct":8.896258388881215},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.046570673584938,"wave_count":1},"fallback_work":185669536,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":432.3626403808594,"n":8,"spread_pct":1.614924526345531},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.151716992259026,"wave_count":1},"fallback_work":185985888,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":430.8921813964844,"n":8,"spread_pct":3.3684255257366154},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":8.57510231435299,"wave_count":1},"fallback_work":185324512,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":432.9901428222656,"n":8,"spread_pct":4.738592667819312},{"cp_plan":{"max_rank_tokens":35328,"summary_ms":6.9689881056547165,"wave_count":1},"fallback_work":1895219104,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":1300.8419189453125,"n":8,"spread_pct":1.7056449225294206},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":10.212128981947899,"wave_count":1},"fallback_work":184635360,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":433.5829772949219,"n":8,"spread_pct":2.32045461152012},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.340684205293655,"wave_count":1},"fallback_work":180272448,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":428.57594299316406,"n":8,"spread_pct":3.835198248822157},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.757678300142288,"wave_count":1},"fallback_work":176944608,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":428.2469482421875,"n":8,"spread_pct":3.0154004946697723},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":9.091373533010483,"wave_count":1},"fallback_work":250820832,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":456.782470703125,"n":8,"spread_pct":4.338800073330515},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":9.90622490644455,"wave_count":1},"fallback_work":235589664,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":456.6753845214844,"n":8,"spread_pct":2.9207370936998953},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":9.76337306201458,"wave_count":1},"fallback_work":225792544,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":441.01734924316406,"n":8,"spread_pct":4.276195637382812},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":9.371235966682434,"wave_count":2},"fallback_work":218956960,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":499.19520568847656,"n":8,"spread_pct":3.446991108766164},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":9.885072708129883,"wave_count":2},"fallback_work":213012768,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":497.3955383300781,"n":8,"spread_pct":3.6484417090316095},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":9.962962940335274,"wave_count":2},"fallback_work":204682080,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":498.61383056640625,"n":8,"spread_pct":3.293928027273855},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":9.918363764882088,"wave_count":2},"fallback_work":199801632,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":498.00555419921875,"n":8,"spread_pct":1.3813646058625209},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":8.221084251999855,"wave_count":2},"fallback_work":192276000,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":496.2950744628906,"n":8,"spread_pct":3.1291435187135237}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.102744489908218,"wave_count":2},"fallback_work":354996544,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":573.9596862792969,"n":8,"spread_pct":9.573164678728622},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":9.387979283928871,"wave_count":1},"fallback_work":679089088,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":679.7449645996094,"n":8,"spread_pct":1.3023324222363923},{"cp_plan":{"max_rank_tokens":20480,"summary_ms":6.65629655122757,"wave_count":1},"fallback_work":1005107936,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":893.7294311523438,"n":8,"spread_pct":1.6915134697360754},{"cp_plan":{"max_rank_tokens":26624,"summary_ms":9.680278599262238,"wave_count":1},"fallback_work":1328826784,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":1126.6359252929688,"n":8,"spread_pct":4.273516543840898},{"cp_plan":{"max_rank_tokens":31232,"summary_ms":7.004978135228157,"wave_count":1},"fallback_work":1651539232,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":1319.6520385742188,"n":8,"spread_pct":36.76081678987071},{"cp_plan":{"max_rank_tokens":36864,"summary_ms":6.832204759120941,"wave_count":1},"fallback_work":1977299392,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":1558.1427612304688,"n":8,"spread_pct":17.402748625845724},{"cp_plan":{"max_rank_tokens":43008,"summary_ms":7.6796021312475204,"wave_count":1},"fallback_work":2306969824,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":1820.8294677734375,"n":8,"spread_pct":1.5511957273400425},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.408359855413437,"wave_count":2},"fallback_work":344646560,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":563.7251281738281,"n":8,"spread_pct":3.6920022011520586},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.489636212587357,"wave_count":2},"fallback_work":351323680,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":568.6424560546875,"n":8,"spread_pct":3.282010098492611},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.270507350564003,"wave_count":2},"fallback_work":344725504,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":561.4661865234375,"n":8,"spread_pct":15.520919415479534},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.351348012685776,"wave_count":2},"fallback_work":344538944,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":565.2498474121094,"n":8,"spread_pct":4.34863053107295},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.619047373533249,"wave_count":2},"fallback_work":345465664,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":572.51416015625,"n":8,"spread_pct":123.86831359076989},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.544506013393402,"wave_count":2},"fallback_work":346673248,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":566.4902648925781,"n":8,"spread_pct":2.9013153383013823},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.476603284478188,"wave_count":2},"fallback_work":347873440,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":571.4268493652344,"n":8,"spread_pct":6.451958835782929},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.349775940179825,"wave_count":2},"fallback_work":344387872,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":561.1487426757812,"n":8,"spread_pct":3.178403502076005},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.48311322927475,"wave_count":2},"fallback_work":347643424,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":572.1498718261719,"n":8,"spread_pct":4.265143861490328},{"cp_plan":{"max_rank_tokens":49280,"summary_ms":7.595304399728775,"wave_count":1},"fallback_work":2636632864,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2067.6622314453125,"n":8,"spread_pct":0.3637203261793443},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.975848063826561,"wave_count":2},"fallback_work":353393824,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":573.3938598632812,"n":8,"spread_pct":5.278735343129506},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.933212116360664,"wave_count":2},"fallback_work":353401216,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":565.2835693359375,"n":8,"spread_pct":3.772196744279556},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.675001069903374,"wave_count":2},"fallback_work":351050944,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":565.8130187988281,"n":8,"spread_pct":4.270345433452634},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.638580769300461,"wave_count":2},"fallback_work":347033056,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":568.5777893066406,"n":8,"spread_pct":3.579552252176287},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.446322426199913,"wave_count":2},"fallback_work":344625280,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":561.0383911132812,"n":8,"spread_pct":7.5861507761106255},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.416477262973785,"wave_count":2},"fallback_work":344776352,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":563.8994140625,"n":8,"spread_pct":2.958405996886219}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.610566258430481,"wave_count":2},"fallback_work":351940000,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":496.1036834716797,"n":8,"spread_pct":2.8070365593996014},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":6.916200742125511,"wave_count":3},"fallback_work":688689728,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":548.6693115234375,"n":8,"spread_pct":4.591675660594848},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":10.31595654785633,"wave_count":1},"fallback_work":1029173984,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":531.5286865234375,"n":8,"spread_pct":3.6115015950961387},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.610486164689064,"wave_count":2},"fallback_work":350929696,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":495.37030029296875,"n":8,"spread_pct":3.5094546710403343},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":6.02419488132,"wave_count":1},"fallback_work":1369519424,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":643.9173583984375,"n":8,"spread_pct":2.4507717672760663},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.566160961985588,"wave_count":2},"fallback_work":349780576,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":495.9710388183594,"n":8,"spread_pct":3.648700259826819},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.532773047685623,"wave_count":2},"fallback_work":350062976,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":500.28424072265625,"n":8,"spread_pct":3.651348348479142}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.4835021942853928,"wave_count":2},"fallback_work":343322496,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":472.359375,"n":8,"spread_pct":6.059022609242168},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":8.602181449532509,"wave_count":3},"fallback_work":682116736,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":527.259033203125,"n":8,"spread_pct":2.3782640081531614}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.409918397665024,"wave_count":2},"fallback_work":348929952,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":467.4011688232422,"n":8,"spread_pct":5.360822938322881}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":8192,"summary_ms":6.547505035996437,"wave_count":1},"fallback_work":596271360,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":477.9779815673828,"n":8,"spread_pct":6.026643775300469},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":3.8505978882312775,"wave_count":2},"fallback_work":466071072,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":531.7486877441406,"n":8,"spread_pct":4.742228127881566},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.7885382771492004,"wave_count":2},"fallback_work":426766176,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":526.1179809570312,"n":8,"spread_pct":3.3171085058011576},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.31857630610466,"wave_count":2},"fallback_work":632951968,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":504.7277069091797,"n":8,"spread_pct":10.581218510824545},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":4.534520208835602,"wave_count":1},"fallback_work":791932832,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":544.4581909179688,"n":8,"spread_pct":3.75292461562555},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":5.121657624840736,"wave_count":2},"fallback_work":1014033824,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":537.7457275390625,"n":8,"spread_pct":2.831281938012986},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":7.972996681928635,"wave_count":1},"fallback_work":1327030432,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":643.1964111328125,"n":8,"spread_pct":2.873878003023307},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":5.873247981071472,"wave_count":3},"fallback_work":450581632,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":593.6418151855469,"n":8,"spread_pct":4.418547684598165},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":10.319141671061516,"wave_count":2},"fallback_work":419234560,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":517.6246948242188,"n":8,"spread_pct":2.5848912415073975}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7680,"summary_ms":6.783602759242058,"wave_count":2},"fallback_work":483744512,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":517.0158996582031,"n":8,"spread_pct":5.2901214568159745},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":5.948048084974289,"wave_count":4},"fallback_work":546113056,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":579.6329956054688,"n":8,"spread_pct":7.077126839861439},{"cp_plan":{"max_rank_tokens":12288,"summary_ms":9.047212079167366,"wave_count":1},"fallback_work":869193024,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":521.6370544433594,"n":8,"spread_pct":4.0904105615613595},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.572192043066025,"wave_count":2},"fallback_work":353601792,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":485.42259216308594,"n":8,"spread_pct":5.16185775754963}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.031230092048645,"wave_count":4},"fallback_work":377277760,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":623.8704528808594,"n":8,"spread_pct":2.594767584600039},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":4.3356530368328094,"wave_count":2},"fallback_work":720117600,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":545.7629699707031,"n":8,"spread_pct":6.7499532112200695},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":8.364200592041016,"wave_count":2},"fallback_work":1056350112,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":620.5447387695312,"n":8,"spread_pct":2.0909992667457136},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.311660632491112,"wave_count":1},"fallback_work":1406084896,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":669.1549377441406,"n":8,"spread_pct":1.4247725017192376},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.665151074528694,"wave_count":2},"fallback_work":372561376,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":494.9322204589844,"n":8,"spread_pct":5.879705183254173},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.169603809714317,"wave_count":2},"fallback_work":367476576,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":496.6262969970703,"n":8,"spread_pct":5.670020626541372},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":14.426665380597115,"wave_count":3},"fallback_work":1755451040,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":786.8639221191406,"n":8,"spread_pct":0.9159338886579745},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.6852844059467316,"wave_count":2},"fallback_work":375893824,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":500.00901794433594,"n":8,"spread_pct":3.5393282549133707},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.197206348180771,"wave_count":2},"fallback_work":367299328,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":493.6322937011719,"n":8,"spread_pct":4.248447776347},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.234099760651588,"wave_count":2},"fallback_work":367696416,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":493.6071319580078,"n":8,"spread_pct":2.8790099686738744}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":4096,"summary_ms":3.738636150956154,"wave_count":1},"fallback_work":263626208,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":434.1714324951172,"n":8,"spread_pct":3.5324909707326517},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.631535917520523,"wave_count":2},"fallback_work":189057248,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":481.4196014404297,"n":8,"spread_pct":7.905803793084158},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.622099757194519,"wave_count":2},"fallback_work":190206368,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":483.2017364501953,"n":8,"spread_pct":2.9543018909620073},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.645366057753563,"wave_count":2},"fallback_work":196244128,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":483.14125061035156,"n":8,"spread_pct":5.667371323979346},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.5970304161310196,"wave_count":2},"fallback_work":198743520,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":485.3108673095703,"n":8,"spread_pct":6.767651198621313},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.621460869908333,"wave_count":2},"fallback_work":200007552,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":487.072998046875,"n":8,"spread_pct":17.725937833168175},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.592001274228096,"wave_count":1},"fallback_work":206504960,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":427.1474151611328,"n":8,"spread_pct":14.23945309082637},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6425013095140457,"wave_count":2},"fallback_work":213715744,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":497.2225646972656,"n":8,"spread_pct":11.344707837217591},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6285463720560074,"wave_count":2},"fallback_work":215899072,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":501.1338806152344,"n":8,"spread_pct":5.958466380714389},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6242101341485977,"wave_count":1},"fallback_work":224062816,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":427.98388671875,"n":8,"spread_pct":5.3144797954617795},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.880208358168602,"wave_count":2},"fallback_work":179690976,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":482.1366882324219,"n":8,"spread_pct":3.3419047380932114},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.625044599175453,"wave_count":1},"fallback_work":226131232,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":425.0187530517578,"n":8,"spread_pct":6.382049792509371},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.866247832775116,"wave_count":2},"fallback_work":180299200,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":483.91064453125,"n":8,"spread_pct":3.9116317659369013},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.7414487451314926,"wave_count":2},"fallback_work":181688128,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":486.5610656738281,"n":8,"spread_pct":11.077324722665153},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.605559468269348,"wave_count":2},"fallback_work":234117504,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":495.3555603027344,"n":8,"spread_pct":3.068640954481296},{"cp_plan":{"max_rank_tokens":4608,"summary_ms":3.879014402627945,"wave_count":3},"fallback_work":387242848,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":528.1023864746094,"n":8,"spread_pct":7.614138895902012},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.090616479516029,"wave_count":1},"fallback_work":508790848,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":464.5553741455078,"n":8,"spread_pct":3.320415984167699},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.44275327026844,"wave_count":1},"fallback_work":637320032,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":470.5565948486328,"n":8,"spread_pct":143.5991825385736},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":4.844453185796738,"wave_count":3},"fallback_work":769928480,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":515.7706909179688,"n":8,"spread_pct":130.976971823894},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":5.036076530814171,"wave_count":1},"fallback_work":904317952,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":503.3880920410156,"n":8,"spread_pct":4.3766297720991085},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.408117547631264,"wave_count":1},"fallback_work":1037213792,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":527.7194213867188,"n":8,"spread_pct":3.46217480486991},{"cp_plan":{"max_rank_tokens":12288,"summary_ms":5.782607942819595,"wave_count":3},"fallback_work":1174993280,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":554.24609375,"n":8,"spread_pct":2.7475759759526945},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":9.805422276258469,"wave_count":3},"fallback_work":1311164000,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":591.3940734863281,"n":8,"spread_pct":2.1434762656368678},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.7121051251888275,"wave_count":3},"fallback_work":1449173312,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":599.8878479003906,"n":8,"spread_pct":5.659026943156744},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":7.023772224783897,"wave_count":1},"fallback_work":1593100480,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":624.929931640625,"n":8,"spread_pct":1.4802245407981287},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.611590713262558,"wave_count":1},"fallback_work":245465120,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":428.08897399902344,"n":8,"spread_pct":4.202052957025246},{"cp_plan":{"max_rank_tokens":16896,"summary_ms":7.485987618565559,"wave_count":1},"fallback_work":1740388992,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":655.2203369140625,"n":8,"spread_pct":0.8608279545753056},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.751387819647789,"wave_count":1},"fallback_work":184738176,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":416.5934295654297,"n":8,"spread_pct":5.969943374202211},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.665562719106674,"wave_count":2},"fallback_work":186639104,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":483.87315368652344,"n":8,"spread_pct":5.066451314178176},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.704102709889412,"wave_count":2},"fallback_work":187156320,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":483.2717590332031,"n":8,"spread_pct":247.73397372000704},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.667309880256653,"wave_count":2},"fallback_work":187472384,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":482.24842834472656,"n":8,"spread_pct":180.3908623143955},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6901049315929413,"wave_count":2},"fallback_work":186811584,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":488.3108215332031,"n":8,"spread_pct":29.147222271351193},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":6.262227892875671,"wave_count":1},"fallback_work":1894490976,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":715.8787841796875,"n":8,"spread_pct":1.5531643987471009},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.906436264514923,"wave_count":1},"fallback_work":186648256,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":416.211669921875,"n":8,"spread_pct":5.770556774008492},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.9820075035095215,"wave_count":2},"fallback_work":182420192,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":480.2757110595703,"n":8,"spread_pct":2.6708139121825147},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.001151770353317,"wave_count":2},"fallback_work":179226560,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":485.3358459472656,"n":8,"spread_pct":5.387001654039979},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":3.83869931101799,"wave_count":1},"fallback_work":251727936,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":438.4090576171875,"n":8,"spread_pct":6.744161893520912},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.715137019753456,"wave_count":2},"fallback_work":236640800,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":495.1288757324219,"n":8,"spread_pct":2.2239181607091005},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.7398729473352432,"wave_count":1},"fallback_work":226983200,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":426.9342956542969,"n":8,"spread_pct":4.939905072151584},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.763822838664055,"wave_count":2},"fallback_work":220284640,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":494.1826171875,"n":8,"spread_pct":5.546064020251244},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.8015712052583694,"wave_count":2},"fallback_work":214476704,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":493.3690185546875,"n":8,"spread_pct":12.31070057765597},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.8271863013505936,"wave_count":2},"fallback_work":206284288,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":496.6857452392578,"n":8,"spread_pct":14.718877697526118},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.991875797510147,"wave_count":2},"fallback_work":201539232,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":485.0405578613281,"n":8,"spread_pct":5.4175559804449644},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.89043428003788,"wave_count":2},"fallback_work":194151200,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":484.60728454589844,"n":8,"spread_pct":6.633040109325093}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.78543883562088,"wave_count":2},"fallback_work":355356896,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":513.0952453613281,"n":8,"spread_pct":2.8351812470094013},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":6.159454584121704,"wave_count":3},"fallback_work":679178912,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":543.0543823242188,"n":8,"spread_pct":8.272031737800345},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":13.865506276488304,"wave_count":2},"fallback_work":1004925760,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":598.6280517578125,"n":8,"spread_pct":3.2863128397636117},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":6.0516782104969025,"wave_count":1},"fallback_work":1328374368,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":653.3472290039062,"n":8,"spread_pct":100.44762729786638},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":11.77191361784935,"wave_count":3},"fallback_work":1650817440,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":761.9842834472656,"n":8,"spread_pct":0.760728401529579},{"cp_plan":{"max_rank_tokens":19456,"summary_ms":8.17088596522808,"wave_count":1},"fallback_work":1976305568,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":851.6564636230469,"n":8,"spread_pct":1.183649496244304},{"cp_plan":{"max_rank_tokens":21504,"summary_ms":9.075583890080452,"wave_count":1},"fallback_work":2305700928,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":960.6603698730469,"n":8,"spread_pct":2.0045097482540615},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.8469117134809494,"wave_count":2},"fallback_work":345540960,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":510.7490539550781,"n":8,"spread_pct":6.08011663323121},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7130918353796005,"wave_count":2},"fallback_work":351818528,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":509.48155212402344,"n":8,"spread_pct":4.704327719982127},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7695616483688354,"wave_count":2},"fallback_work":345488384,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":509.89320373535156,"n":8,"spread_pct":4.685321660249002},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.27629628777504,"wave_count":2},"fallback_work":345564640,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":514.7276916503906,"n":8,"spread_pct":147.8954349831208},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7327054888010025,"wave_count":2},"fallback_work":346096608,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":514.4635314941406,"n":8,"spread_pct":132.30070030497794},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.637490957975388,"wave_count":2},"fallback_work":347303296,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":513.2456359863281,"n":8,"spread_pct":3.1930904166973626},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.606535658240318,"wave_count":2},"fallback_work":348371168,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":513.7791442871094,"n":8,"spread_pct":3.3752338434553386},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7966854870319366,"wave_count":2},"fallback_work":345282464,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":520.8036804199219,"n":8,"spread_pct":208.57639258172605},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.597112536430359,"wave_count":2},"fallback_work":348141344,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":512.8129272460938,"n":8,"spread_pct":7.524648809731387},{"cp_plan":{"max_rank_tokens":24576,"summary_ms":8.5055623203516,"wave_count":1},"fallback_work":2634957472,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1069.5607299804688,"n":8,"spread_pct":1.420306450419127},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7864521145820618,"wave_count":2},"fallback_work":353886944,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":511.4193572998047,"n":8,"spread_pct":1.4569413957277426},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7463195621967316,"wave_count":2},"fallback_work":354025536,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":512.7771148681641,"n":8,"spread_pct":3.04856121021881},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.742788150906563,"wave_count":2},"fallback_work":351808544,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":528.4214477539062,"n":8,"spread_pct":8.089255557248341},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.764126613736153,"wave_count":2},"fallback_work":347925440,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":516.5192565917969,"n":8,"spread_pct":137.42347182794896},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.8369055837392807,"wave_count":2},"fallback_work":345650880,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":515.5220336914062,"n":8,"spread_pct":8.248196579975192},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.878546878695488,"wave_count":2},"fallback_work":345933280,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":513.4075622558594,"n":8,"spread_pct":4.415835643566358}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12204,"summary_ms":0.0004637986421585083,"wave_count":0},"fallback_work":352228864,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":406.22784423828125,"n":8,"spread_pct":6.5525358429253275},{"cp_plan":{"max_rank_tokens":23926,"summary_ms":0.0002942979335784912,"wave_count":0},"fallback_work":689822464,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":666.2142639160156,"n":8,"spread_pct":16.331813426796415},{"cp_plan":{"max_rank_tokens":35778,"summary_ms":0.00041350722312927246,"wave_count":0},"fallback_work":1031160064,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":948.2279052734375,"n":8,"spread_pct":99.45944087983818},{"cp_plan":{"max_rank_tokens":12154,"summary_ms":0.0003166496753692627,"wave_count":0},"fallback_work":351018240,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":401.5489044189453,"n":8,"spread_pct":5.395952691582875},{"cp_plan":{"max_rank_tokens":47640,"summary_ms":0.00038929283618927,"wave_count":0},"fallback_work":1372556288,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1245.9404907226562,"n":8,"spread_pct":1.2376130312657478},{"cp_plan":{"max_rank_tokens":12114,"summary_ms":0.00037997961044311523,"wave_count":0},"fallback_work":349866240,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":405.5564727783203,"n":8,"spread_pct":8.017661159757065},{"cp_plan":{"max_rank_tokens":12109,"summary_ms":0.000346451997756958,"wave_count":0},"fallback_work":349951616,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":403.44525146484375,"n":8,"spread_pct":12.971208828015898}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11920,"summary_ms":0.00045821070671081543,"wave_count":0},"fallback_work":343787520,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":364.3822784423828,"n":8,"spread_pct":3.9023696535998487},{"cp_plan":{"max_rank_tokens":23728,"summary_ms":0.000391155481338501,"wave_count":0},"fallback_work":683628544,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":628.7559814453125,"n":8,"spread_pct":1.1135230486942824}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12138,"summary_ms":0.0003371387720108032,"wave_count":0},"fallback_work":349705472,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":367.83148193359375,"n":8,"spread_pct":4.121445251189279}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":20701,"summary_ms":0.0003334134817123413,"wave_count":0},"fallback_work":597073536,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":595.2452697753906,"n":8,"spread_pct":112.2725012467318},{"cp_plan":{"max_rank_tokens":16154,"summary_ms":0.0003241002559661865,"wave_count":0},"fallback_work":466349312,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":479.47216796875,"n":8,"spread_pct":0.7961126686552087},{"cp_plan":{"max_rank_tokens":14771,"summary_ms":0.00027567148208618164,"wave_count":0},"fallback_work":426748288,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":445.2295379638672,"n":8,"spread_pct":5.333264698470586},{"cp_plan":{"max_rank_tokens":21963,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":633648512,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":603.8655090332031,"n":8,"spread_pct":2.99164142551863},{"cp_plan":{"max_rank_tokens":27497,"summary_ms":0.00024214386940002441,"wave_count":0},"fallback_work":793027712,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":716.4458618164062,"n":8,"spread_pct":0.6431112510903457},{"cp_plan":{"max_rank_tokens":35243,"summary_ms":0.00023096799850463867,"wave_count":0},"fallback_work":1015883136,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":884.6344299316406,"n":8,"spread_pct":0.7519671877393795},{"cp_plan":{"max_rank_tokens":46153,"summary_ms":0.00034831464290618896,"wave_count":0},"fallback_work":1329861760,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":1161.292236328125,"n":8,"spread_pct":0.5532356469980046},{"cp_plan":{"max_rank_tokens":15600,"summary_ms":0.0002868473529815674,"wave_count":0},"fallback_work":450623488,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":462.7440643310547,"n":8,"spread_pct":3.643314676868121},{"cp_plan":{"max_rank_tokens":14494,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":419000064,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":442.5244445800781,"n":8,"spread_pct":3.069085014931409}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16800,"summary_ms":0.0004153698682785034,"wave_count":0},"fallback_work":484462592,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":484.4833679199219,"n":8,"spread_pct":2.2671906001876723},{"cp_plan":{"max_rank_tokens":18971,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":546987392,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":520.5426025390625,"n":8,"spread_pct":2.3310238410425193},{"cp_plan":{"max_rank_tokens":30232,"summary_ms":0.00033155083656311035,"wave_count":0},"fallback_work":871074816,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":786.337890625,"n":8,"spread_pct":0.8059776331640194},{"cp_plan":{"max_rank_tokens":12255,"summary_ms":0.0003166496753692627,"wave_count":0},"fallback_work":353795968,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":382.0907897949219,"n":8,"spread_pct":487.3759769194969}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13078,"summary_ms":0.00022351741790771484,"wave_count":0},"fallback_work":377531136,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":434.26025390625,"n":8,"spread_pct":4.598287823334626},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":721230336,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":710.7235107421875,"n":8,"spread_pct":0.7824470517997605},{"cp_plan":{"max_rank_tokens":36716,"summary_ms":0.00019744038581848145,"wave_count":0},"fallback_work":1058305536,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":975.8497924804688,"n":8,"spread_pct":0.714702849073039},{"cp_plan":{"max_rank_tokens":48890,"summary_ms":0.0001862645149230957,"wave_count":0},"fallback_work":1408916736,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":1267.6852416992188,"n":8,"spread_pct":0.7209040260706012},{"cp_plan":{"max_rank_tokens":12899,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":372605312,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":428.94158935546875,"n":8,"spread_pct":3.6802523243829954},{"cp_plan":{"max_rank_tokens":12722,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":367507712,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":425.0924530029297,"n":8,"spread_pct":4.144275973398623},{"cp_plan":{"max_rank_tokens":61066,"summary_ms":0.0002998858690261841,"wave_count":0},"fallback_work":1759356160,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1602.3207397460938,"n":8,"spread_pct":0.6106636328387669},{"cp_plan":{"max_rank_tokens":13015,"summary_ms":0.00022351741790771484,"wave_count":0},"fallback_work":375946112,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":427.3132629394531,"n":8,"spread_pct":3.2028893171866337},{"cp_plan":{"max_rank_tokens":12701,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":367132288,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":422.1684265136719,"n":8,"spread_pct":3.6965970210967996},{"cp_plan":{"max_rank_tokens":12700,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":367332864,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":420.2462463378906,"n":8,"spread_pct":2.7789507654298147}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9058,"summary_ms":0.00029616057872772217,"wave_count":0},"fallback_work":262803712,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":380.8225555419922,"n":8,"spread_pct":7.274966125531286},{"cp_plan":{"max_rank_tokens":6403,"summary_ms":0.00020116567611694336,"wave_count":0},"fallback_work":187257216,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":375.4839172363281,"n":8,"spread_pct":4.210663371656547},{"cp_plan":{"max_rank_tokens":6443,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":188409216,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":373.0928497314453,"n":8,"spread_pct":9.67146841727989},{"cp_plan":{"max_rank_tokens":6668,"summary_ms":0.0001862645149230957,"wave_count":0},"fallback_work":194659840,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":373.0132751464844,"n":8,"spread_pct":4.306997873417755},{"cp_plan":{"max_rank_tokens":6755,"summary_ms":0.00023283064365386963,"wave_count":0},"fallback_work":197165440,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":374.4326629638672,"n":8,"spread_pct":2.172353934392686},{"cp_plan":{"max_rank_tokens":6799,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":198432640,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":373.5226135253906,"n":8,"spread_pct":5.561461254949002},{"cp_plan":{"max_rank_tokens":7040,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":205144064,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":375.0883483886719,"n":8,"spread_pct":3.068027185977913},{"cp_plan":{"max_rank_tokens":7291,"summary_ms":0.00021047890186309814,"wave_count":0},"fallback_work":212372864,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":375.37744140625,"n":8,"spread_pct":4.002433097936582},{"cp_plan":{"max_rank_tokens":7367,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":214561664,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":375.18751525878906,"n":8,"spread_pct":4.928647129533902},{"cp_plan":{"max_rank_tokens":7666,"summary_ms":0.00023283064365386963,"wave_count":0},"fallback_work":222943488,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":376.1680145263672,"n":8,"spread_pct":5.688458884187504},{"cp_plan":{"max_rank_tokens":5988,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":176681472,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":371.89984130859375,"n":8,"spread_pct":6.09544366373898},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":225017088,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":380.967529296875,"n":8,"spread_pct":10.590462131586992},{"cp_plan":{"max_rank_tokens":6024,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":177488896,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":372.8534698486328,"n":8,"spread_pct":5.7207751482740665},{"cp_plan":{"max_rank_tokens":6102,"summary_ms":0.0002216547727584839,"wave_count":0},"fallback_work":179276544,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":372.8459777832031,"n":8,"spread_pct":3.6229180030443424},{"cp_plan":{"max_rank_tokens":8016,"summary_ms":0.00021047890186309814,"wave_count":0},"fallback_work":233023488,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":379.55921936035156,"n":8,"spread_pct":4.735749961818778},{"cp_plan":{"max_rank_tokens":13361,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":386730112,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":420.09974670410156,"n":8,"spread_pct":10.51154313015014},{"cp_plan":{"max_rank_tokens":17592,"summary_ms":0.00023096799850463867,"wave_count":0},"fallback_work":508582912,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":489.66578674316406,"n":8,"spread_pct":2.6240843268884726},{"cp_plan":{"max_rank_tokens":22066,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":637434112,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":582.5015563964844,"n":8,"spread_pct":1.3655911742037723},{"cp_plan":{"max_rank_tokens":26682,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":770374912,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":680.5422668457031,"n":8,"spread_pct":0.8151297158669308},{"cp_plan":{"max_rank_tokens":31360,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":905101312,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":764.0555725097656,"n":8,"spread_pct":0.8774044275148999},{"cp_plan":{"max_rank_tokens":35986,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":1038330112,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":867.4228515625,"n":8,"spread_pct":0.35323324413081586},{"cp_plan":{"max_rank_tokens":40782,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":1176454912,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":965.6492309570312,"n":8,"spread_pct":0.6553676058771027},{"cp_plan":{"max_rank_tokens":45522,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":1312966912,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":1066.74365234375,"n":8,"spread_pct":0.8852167194353802},{"cp_plan":{"max_rank_tokens":50326,"summary_ms":0.00024400651454925537,"wave_count":0},"fallback_work":1451322112,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":1174.8190307617188,"n":8,"spread_pct":0.8202933931344873},{"cp_plan":{"max_rank_tokens":55336,"summary_ms":0.00023655593395233154,"wave_count":0},"fallback_work":1595610112,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":1260.443115234375,"n":8,"spread_pct":0.7906891158340235},{"cp_plan":{"max_rank_tokens":8411,"summary_ms":0.000303611159324646,"wave_count":0},"fallback_work":244399488,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":378.55677795410156,"n":8,"spread_pct":5.740434647822934},{"cp_plan":{"max_rank_tokens":60463,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":1743267712,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":1366.2979125976562,"n":8,"spread_pct":0.7765677890859789},{"cp_plan":{"max_rank_tokens":6223,"summary_ms":0.0002477318048477173,"wave_count":0},"fallback_work":182531968,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":372.8401184082031,"n":8,"spread_pct":3.167331370257976},{"cp_plan":{"max_rank_tokens":6304,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":184635392,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":374.1065979003906,"n":8,"spread_pct":3.0582054822177676},{"cp_plan":{"max_rank_tokens":6322,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":185153792,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":375.2043151855469,"n":8,"spread_pct":10.284793563402753},{"cp_plan":{"max_rank_tokens":6333,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":185470592,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":377.36126708984375,"n":8,"spread_pct":3.3542131438968203},{"cp_plan":{"max_rank_tokens":6310,"summary_ms":0.00025331974029541016,"wave_count":0},"fallback_work":184808192,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":374.2309112548828,"n":8,"spread_pct":4.382849477629434},{"cp_plan":{"max_rank_tokens":65842,"summary_ms":0.00042654573917388916,"wave_count":0},"fallback_work":1897953536,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":1525.9022216796875,"n":8,"spread_pct":0.7262849023666919},{"cp_plan":{"max_rank_tokens":6245,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":183853696,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":376.21067810058594,"n":8,"spread_pct":167.945446956467},{"cp_plan":{"max_rank_tokens":6083,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":179417472,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":371.9134216308594,"n":8,"spread_pct":3.6049649872867158},{"cp_plan":{"max_rank_tokens":5957,"summary_ms":0.00021047890186309814,"wave_count":0},"fallback_work":176018048,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":371.9379119873047,"n":8,"spread_pct":3.66386939775116},{"cp_plan":{"max_rank_tokens":8629,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":250677888,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":378.6429443359375,"n":8,"spread_pct":7.972550512711314},{"cp_plan":{"max_rank_tokens":8089,"summary_ms":0.0005457550287246704,"wave_count":0},"fallback_work":235355264,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":377.21685791015625,"n":8,"spread_pct":7.1347184070212535},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":225475840,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":376.7103729248047,"n":8,"spread_pct":7.933800956906713},{"cp_plan":{"max_rank_tokens":7490,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":218562816,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":374.4869842529297,"n":8,"spread_pct":8.431057152201912},{"cp_plan":{"max_rank_tokens":7273,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":212542592,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":373.51275634765625,"n":8,"spread_pct":7.301017887837391},{"cp_plan":{"max_rank_tokens":6973,"summary_ms":0.00019744038581848145,"wave_count":0},"fallback_work":204131968,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":374.66563415527344,"n":8,"spread_pct":4.999884540606214},{"cp_plan":{"max_rank_tokens":6793,"summary_ms":0.00019744038581848145,"wave_count":0},"fallback_work":199177344,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":378.9352569580078,"n":8,"spread_pct":4.1490299892876275},{"cp_plan":{"max_rank_tokens":6521,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":191573120,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":375.51544189453125,"n":8,"spread_pct":5.3267197024922615}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12291,"summary_ms":0.0002514570951461792,"wave_count":0},"fallback_work":355258752,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":438.7007293701172,"n":8,"spread_pct":3.03933564588065},{"cp_plan":{"max_rank_tokens":23563,"summary_ms":0.00022351741790771484,"wave_count":0},"fallback_work":679892352,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":689.674560546875,"n":8,"spread_pct":0.6125157394562016},{"cp_plan":{"max_rank_tokens":34902,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":1006455552,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":954.8472900390625,"n":8,"spread_pct":1.0133967706151883},{"cp_plan":{"max_rank_tokens":46161,"summary_ms":0.0002216547727584839,"wave_count":0},"fallback_work":1330714752,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":1222.7223510742188,"n":8,"spread_pct":0.9168134530266759},{"cp_plan":{"max_rank_tokens":57385,"summary_ms":0.00021047890186309814,"wave_count":0},"fallback_work":1653965952,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":1492.2111206054688,"n":8,"spread_pct":0.3781762780505724},{"cp_plan":{"max_rank_tokens":68715,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":1980269952,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":1772.3653564453125,"n":8,"spread_pct":0.9339002234073641},{"cp_plan":{"max_rank_tokens":80181,"summary_ms":0.00034831464290618896,"wave_count":0},"fallback_work":2310490752,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":2040.830810546875,"n":8,"spread_pct":0.5277571186768946},{"cp_plan":{"max_rank_tokens":11890,"summary_ms":0.00019744038581848145,"wave_count":0},"fallback_work":344627456,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":429.1795654296875,"n":8,"spread_pct":2.2261116052242436},{"cp_plan":{"max_rank_tokens":12153,"summary_ms":0.00019744038581848145,"wave_count":0},"fallback_work":351513728,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":432.8504638671875,"n":8,"spread_pct":2.5562471816128496},{"cp_plan":{"max_rank_tokens":11903,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":344772480,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":429.5210876464844,"n":8,"spread_pct":2.187196504441671},{"cp_plan":{"max_rank_tokens":11876,"summary_ms":0.00027567148208618164,"wave_count":0},"fallback_work":344453632,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":431.1886749267578,"n":8,"spread_pct":147.724542322506},{"cp_plan":{"max_rank_tokens":11939,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":345579904,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":431.9964141845703,"n":8,"spread_pct":3.0305407551333623},{"cp_plan":{"max_rank_tokens":11981,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":346789504,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":431.98692321777344,"n":8,"spread_pct":149.0570189110511},{"cp_plan":{"max_rank_tokens":12033,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":348057728,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":434.5507354736328,"n":8,"spread_pct":2.526897542470049},{"cp_plan":{"max_rank_tokens":11881,"summary_ms":0.0002384185791015625,"wave_count":0},"fallback_work":344368256,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":428.485107421875,"n":8,"spread_pct":3.0337236300443},{"cp_plan":{"max_rank_tokens":12025,"summary_ms":0.0003259629011154175,"wave_count":0},"fallback_work":347827328,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":430.1382293701172,"n":8,"spread_pct":1.787975319497066},{"cp_plan":{"max_rank_tokens":91657,"summary_ms":0.00038370490074157715,"wave_count":0},"fallback_work":2640770176,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2362.8453369140625,"n":8,"spread_pct":0.38681721716335177},{"cp_plan":{"max_rank_tokens":12225,"summary_ms":0.000303611159324646,"wave_count":0},"fallback_work":353587328,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":437.172119140625,"n":8,"spread_pct":3.684492619189493},{"cp_plan":{"max_rank_tokens":12215,"summary_ms":0.00024959444999694824,"wave_count":0},"fallback_work":353528704,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":434.55145263671875,"n":8,"spread_pct":3.176135031702921},{"cp_plan":{"max_rank_tokens":12123,"summary_ms":0.00023469328880310059,"wave_count":0},"fallback_work":351108480,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":433.25807189941406,"n":8,"spread_pct":2.6392335479596554},{"cp_plan":{"max_rank_tokens":11973,"summary_ms":0.00023469328880310059,"wave_count":0},"fallback_work":347017856,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":433.948486328125,"n":8,"spread_pct":4.780527248821769},{"cp_plan":{"max_rank_tokens":11879,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":344540032,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":430.217041015625,"n":8,"spread_pct":2.7234689181344125},{"cp_plan":{"max_rank_tokens":11874,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":344625408,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":431.5845184326172,"n":8,"spread_pct":3.9686012636752874}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":74719,"summary_ms":0.0005606561899185181,"wave_count":0},"fallback_work":2156330880,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":2240.1514892578125,"n":8,"spread_pct":0.7161337322800904},{"cp_plan":{"max_rank_tokens":14302,"summary_ms":0.0002980232238769531,"wave_count":0},"fallback_work":416550656,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":496.9385070800781,"n":8,"spread_pct":0.6333764460259323},{"cp_plan":{"max_rank_tokens":106432,"summary_ms":0.0010617077350616455,"wave_count":0},"fallback_work":3069435904,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":2993.8665771484375,"n":8,"spread_pct":0.4011619621903587},{"cp_plan":{"max_rank_tokens":13279,"summary_ms":0.0003725290298461914,"wave_count":0},"fallback_work":387317632,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":473.1605987548828,"n":8,"spread_pct":0.6710749921966337}],"cell":"cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":38879,"summary_ms":10.211452841758728,"wave_count":1},"fallback_work":2153858560,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":1447.8024291992188,"n":8,"spread_pct":0.9834156064330897},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":11.90943457186222,"wave_count":1},"fallback_work":417043808,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":446.35919189453125,"n":8,"spread_pct":1.575579822043828},{"cp_plan":{"max_rank_tokens":55296,"summary_ms":7.7382829040288925,"wave_count":1},"fallback_work":3065375744,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":1827.8302612304688,"n":8,"spread_pct":0.7016956626721101},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":30.231239274144173,"wave_count":1},"fallback_work":387925504,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":443.5535583496094,"n":8,"spread_pct":2.1577072211558224}],"cell":"cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19456,"summary_ms":7.792418822646141,"wave_count":1},"fallback_work":2154293504,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":753.9234619140625,"n":8,"spread_pct":0.36317985615937204},{"cp_plan":{"max_rank_tokens":4608,"summary_ms":4.450755193829536,"wave_count":3},"fallback_work":419059968,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":480.3867950439453,"n":8,"spread_pct":4.60861130752454},{"cp_plan":{"max_rank_tokens":26880,"summary_ms":8.49173218011856,"wave_count":1},"fallback_work":3064918528,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":930.5277404785156,"n":8,"spread_pct":0.5922955726758639},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":4.488008096814156,"wave_count":1},"fallback_work":390097152,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":420.1151885986328,"n":8,"spread_pct":6.711227306005095}],"cell":"cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":74719,"summary_ms":0.00040605664253234863,"wave_count":0},"fallback_work":2156330880,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":1565.0365600585938,"n":8,"spread_pct":87.8281854859137},{"cp_plan":{"max_rank_tokens":14302,"summary_ms":0.00033527612686157227,"wave_count":0},"fallback_work":416550656,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":400.2593994140625,"n":8,"spread_pct":3.143324412726732},{"cp_plan":{"max_rank_tokens":106432,"summary_ms":0.0005774199962615967,"wave_count":0},"fallback_work":3069435904,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":2028.8363037109375,"n":8,"spread_pct":0.601165074339839},{"cp_plan":{"max_rank_tokens":13279,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":387317632,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":389.78773498535156,"n":8,"spread_pct":2.815143348855275}],"cell":"cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":58351,"summary_ms":0.0006146728992462158,"wave_count":0},"fallback_work":1682835328,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1820.9386596679688,"n":8,"spread_pct":0.2644744754679511},{"cp_plan":{"max_rank_tokens":24556,"summary_ms":0.0006277114152908325,"wave_count":0},"fallback_work":710227456,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":774.426025390625,"n":8,"spread_pct":1.1359376857040986},{"cp_plan":{"max_rank_tokens":73696,"summary_ms":0.0010021030902862549,"wave_count":0},"fallback_work":2124541952,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":2124.5123291015625,"n":8,"spread_pct":0.5533668830823709},{"cp_plan":{"max_rank_tokens":21487,"summary_ms":0.00040046870708465576,"wave_count":0},"fallback_work":622069632,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":718.0171203613281,"n":8,"spread_pct":0.8314779187422978}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":31727,"summary_ms":10.47857478260994,"wave_count":1},"fallback_work":1680624384,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1213.57373046875,"n":8,"spread_pct":0.7784967323154149},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":8.550979197025299,"wave_count":1},"fallback_work":709835200,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":569.1647033691406,"n":8,"spread_pct":1.8339671092115022},{"cp_plan":{"max_rank_tokens":37888,"summary_ms":7.6635610312223434,"wave_count":1},"fallback_work":2121528832,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":1286.4663696289062,"n":8,"spread_pct":0.9641861580602792},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":10.286606848239899,"wave_count":1},"fallback_work":621890304,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":558.6167297363281,"n":8,"spread_pct":84.15418662027288}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":15360,"summary_ms":7.34776072204113,"wave_count":1},"fallback_work":1680403584,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":633.5895385742188,"n":8,"spread_pct":2.4496637086540436},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.592370241880417,"wave_count":1},"fallback_work":710818720,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":463.6352081298828,"n":8,"spread_pct":4.4533521155333045},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":6.912071257829666,"wave_count":1},"fallback_work":2120808704,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":703.4944152832031,"n":8,"spread_pct":2.2342167884451336},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.591016098856926,"wave_count":1},"fallback_work":623078528,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":445.5564727783203,"n":8,"spread_pct":2.648822915652468}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":58351,"summary_ms":0.0004991888999938965,"wave_count":0},"fallback_work":1682835328,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1252.2802734375,"n":8,"spread_pct":0.31843278082560333},{"cp_plan":{"max_rank_tokens":24556,"summary_ms":0.0002514570951461792,"wave_count":0},"fallback_work":710227456,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":560.3367614746094,"n":8,"spread_pct":0.5102522368603432},{"cp_plan":{"max_rank_tokens":73696,"summary_ms":0.00039301812648773193,"wave_count":0},"fallback_work":2124541952,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":1449.23046875,"n":8,"spread_pct":1.2084976401808072},{"cp_plan":{"max_rank_tokens":21487,"summary_ms":0.0003948807716369629,"wave_count":0},"fallback_work":622069632,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":522.3816223144531,"n":8,"spread_pct":0.7796978057591805}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":37871,"summary_ms":0.001130625605583191,"wave_count":0},"fallback_work":1093011328,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1137.8681640625,"n":8,"spread_pct":1.0004917707781735},{"cp_plan":{"max_rank_tokens":10222,"summary_ms":0.0003241002559661865,"wave_count":0},"fallback_work":296949504,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":339.41111755371094,"n":8,"spread_pct":1.1377900395560134},{"cp_plan":{"max_rank_tokens":53216,"summary_ms":0.0013150274753570557,"wave_count":0},"fallback_work":1534717952,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":1477.4332885742188,"n":8,"spread_pct":1.093485418466218},{"cp_plan":{"max_rank_tokens":9199,"summary_ms":0.0006966292858123779,"wave_count":0},"fallback_work":267716480,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":322.16636657714844,"n":8,"spread_pct":1.262794246049742}],"cell":"cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":21487,"summary_ms":75.50043985247612,"wave_count":1},"fallback_work":1091783424,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":817.171875,"n":8,"spread_pct":1.251151730912637},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":7.29801133275032,"wave_count":1},"fallback_work":297114208,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":398.825927734375,"n":8,"spread_pct":2.348935565879097},{"cp_plan":{"max_rank_tokens":28672,"summary_ms":74.11016896367073,"wave_count":1},"fallback_work":1532687872,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":931.36572265625,"n":8,"spread_pct":1.1892987815587093},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":12.122849002480507,"wave_count":1},"fallback_work":267995904,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":400.0583801269531,"n":8,"spread_pct":3.049481684992722}],"cell":"cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.699783563613892,"wave_count":1},"fallback_work":1092054144,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":482.8745422363281,"n":8,"spread_pct":4.255135600559015},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.802662715315819,"wave_count":3},"fallback_work":298179712,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":469.7376708984375,"n":8,"spread_pct":6.114005619929467},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":6.775403395295143,"wave_count":1},"fallback_work":1532459264,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":547.840087890625,"n":8,"spread_pct":2.3066325364212172},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.9200540632009506,"wave_count":1},"fallback_work":269216896,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":400.803466796875,"n":8,"spread_pct":6.011408358820265}],"cell":"cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":37871,"summary_ms":0.0005383044481277466,"wave_count":0},"fallback_work":1093011328,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":811.1132202148438,"n":8,"spread_pct":0.45208622679981364},{"cp_plan":{"max_rank_tokens":10222,"summary_ms":0.0003203749656677246,"wave_count":0},"fallback_work":296949504,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":338.5749969482422,"n":8,"spread_pct":5.887091675003812},{"cp_plan":{"max_rank_tokens":53216,"summary_ms":0.0008344650268554688,"wave_count":0},"fallback_work":1534717952,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":1032.671875,"n":8,"spread_pct":0.7550314150186863},{"cp_plan":{"max_rank_tokens":9199,"summary_ms":0.000391155481338501,"wave_count":0},"fallback_work":267716480,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":334.4553527832031,"n":8,"spread_pct":6.28181220689063}],"cell":"cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":28852,"summary_ms":0.0004600733518600464,"wave_count":0},"fallback_work":833952256,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":888.7629699707031,"n":8,"spread_pct":1.1558705263213918},{"cp_plan":{"max_rank_tokens":33460,"summary_ms":0.0003110617399215698,"wave_count":0},"fallback_work":966203904,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":1003.4216918945312,"n":8,"spread_pct":0.3936846706347379},{"cp_plan":{"max_rank_tokens":29364,"summary_ms":0.0003203749656677246,"wave_count":0},"fallback_work":848468480,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":902.4606628417969,"n":8,"spread_pct":1.1217770046095557},{"cp_plan":{"max_rank_tokens":36532,"summary_ms":0.00019744038581848145,"wave_count":0},"fallback_work":1054448128,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":1081.2529907226562,"n":8,"spread_pct":0.6535499444574558},{"cp_plan":{"max_rank_tokens":48820,"summary_ms":0.0010542571544647217,"wave_count":0},"fallback_work":1408113152,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":1481.078125,"n":8,"spread_pct":0.656276704047938}],"cell":"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":18432,"summary_ms":5.49769401550293,"wave_count":1},"fallback_work":833353792,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":653.9557189941406,"n":8,"spread_pct":1.1178029813212775},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":4.106847569346428,"wave_count":1},"fallback_work":965253184,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":761.0550231933594,"n":8,"spread_pct":1.8476355859919766},{"cp_plan":{"max_rank_tokens":18944,"summary_ms":5.482662469148636,"wave_count":1},"fallback_work":847779904,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":661.6395263671875,"n":8,"spread_pct":2.392885364263837},{"cp_plan":{"max_rank_tokens":25088,"summary_ms":5.043867975473404,"wave_count":1},"fallback_work":1053284416,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":830.4454040527344,"n":8,"spread_pct":1.6357541800898021},{"cp_plan":{"max_rank_tokens":30720,"summary_ms":4.913879558444023,"wave_count":1},"fallback_work":1406294080,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":1023.1465148925781,"n":8,"spread_pct":114.63374381999414}],"cell":"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13312,"summary_ms":12.120414525270462,"wave_count":3},"fallback_work":834234208,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":543.7544555664062,"n":8,"spread_pct":2.369109987312649},{"cp_plan":{"max_rank_tokens":16384,"summary_ms":5.290728062391281,"wave_count":2},"fallback_work":965760864,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":559.3002014160156,"n":8,"spread_pct":0.6313475161903251},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":13.749759644269943,"wave_count":3},"fallback_work":848516960,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":546.0130004882812,"n":8,"spread_pct":6.342273652991206},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":5.629280582070351,"wave_count":2},"fallback_work":1053587296,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":600.7764282226562,"n":8,"spread_pct":0.6573220018325489},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":6.452424451708794,"wave_count":1},"fallback_work":1406170976,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":731.8774719238281,"n":8,"spread_pct":0.9295242409556003}],"cell":"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":28852,"summary_ms":0.0003129243850708008,"wave_count":0},"fallback_work":833952256,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":652.8428039550781,"n":8,"spread_pct":0.910334641542237},{"cp_plan":{"max_rank_tokens":33460,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":966203904,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":729.4722290039062,"n":8,"spread_pct":1.093997467969547},{"cp_plan":{"max_rank_tokens":29364,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":848468480,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":659.9429626464844,"n":8,"spread_pct":0.6703996589874567},{"cp_plan":{"max_rank_tokens":36532,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":1054448128,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":779.5125427246094,"n":8,"spread_pct":1.022296970054536},{"cp_plan":{"max_rank_tokens":48820,"summary_ms":0.0004600733518600464,"wave_count":0},"fallback_work":1408113152,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":1040.0916137695312,"n":8,"spread_pct":0.4640258493981026}],"cell":"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16756,"summary_ms":0.000303611159324646,"wave_count":0},"fallback_work":484473344,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":518.5059509277344,"n":8,"spread_pct":0.988357957136155},{"cp_plan":{"max_rank_tokens":17524,"summary_ms":0.00037439167499542236,"wave_count":0},"fallback_work":506362368,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":532.5019226074219,"n":8,"spread_pct":1.1604890428535741},{"cp_plan":{"max_rank_tokens":25204,"summary_ms":0.0009033828973770142,"wave_count":0},"fallback_work":727316992,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":753.3235168457031,"n":8,"spread_pct":0.8317390729273618}],"cell":"cal-hetero3|Qwen/Qwen3-1.7B|L28|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9588,"summary_ms":5.157250910997391,"wave_count":1},"fallback_work":484160576,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":416.4290771484375,"n":8,"spread_pct":3.423233716960286},{"cp_plan":{"max_rank_tokens":10356,"summary_ms":5.273092538118362,"wave_count":1},"fallback_work":505947200,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":427.51983642578125,"n":8,"spread_pct":1.6788529063878357},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":6.3834842294454575,"wave_count":1},"fallback_work":726467648,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":507.2281188964844,"n":8,"spread_pct":3.006362309607372}],"cell":"cal-hetero3|Qwen/Qwen3-1.7B|L28|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":10.0253876298666,"wave_count":3},"fallback_work":484741472,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":420.36158752441406,"n":8,"spread_pct":2.4897185240295436},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":9.805496782064438,"wave_count":3},"fallback_work":506378592,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":425.19940185546875,"n":8,"spread_pct":4.41199965491834},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":4.474017769098282,"wave_count":3},"fallback_work":726583648,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":458.00502014160156,"n":8,"spread_pct":4.23737258259537}],"cell":"cal-hetero3|Qwen/Qwen3-1.7B|L28|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16756,"summary_ms":0.0002868473529815674,"wave_count":0},"fallback_work":484473344,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":402.9604949951172,"n":8,"spread_pct":4.112673310734731},{"cp_plan":{"max_rank_tokens":17524,"summary_ms":0.00030919909477233887,"wave_count":0},"fallback_work":506362368,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":415.33399963378906,"n":8,"spread_pct":0.9199198113957916},{"cp_plan":{"max_rank_tokens":25204,"summary_ms":0.0003762543201446533,"wave_count":0},"fallback_work":727316992,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":561.1549072265625,"n":8,"spread_pct":0.8768145876609018}],"cell":"cal-hetero3|Qwen/Qwen3-1.7B|L28|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19380,"summary_ms":0.0003948807716369629,"wave_count":0},"fallback_work":560404992,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":585.1304626464844,"n":8,"spread_pct":0.9736014004705961},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":722147840,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":722.0768432617188,"n":8,"spread_pct":0.8497860658042804},{"cp_plan":{"max_rank_tokens":20916,"summary_ms":0.0003166496753692627,"wave_count":0},"fallback_work":604412416,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":615.6943359375,"n":8,"spread_pct":1.753916912911141},{"cp_plan":{"max_rank_tokens":29108,"summary_ms":0.0008158385753631592,"wave_count":0},"fallback_work":839883264,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":859.3929443359375,"n":8,"spread_pct":1.5710041281808462}],"cell":"cal-hetero|Qwen/Qwen3-1.7B|L28|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.830468773841858,"wave_count":1},"fallback_work":560064576,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":462.65386962890625,"n":8,"spread_pct":1.477905168004574},{"cp_plan":{"max_rank_tokens":15872,"summary_ms":5.573596805334091,"wave_count":1},"fallback_work":721406016,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":545.3937683105469,"n":8,"spread_pct":1.5110570410491855},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":7.491635158658028,"wave_count":1},"fallback_work":603932736,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":483.70628356933594,"n":8,"spread_pct":1.364099725001518},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":7.542785257101059,"wave_count":1},"fallback_work":838879296,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":593.2567749023438,"n":8,"spread_pct":1.2164298900021697}],"cell":"cal-hetero|Qwen/Qwen3-1.7B|L28|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6580,"summary_ms":4.180407151579857,"wave_count":3},"fallback_work":560779104,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":449.16297912597656,"n":8,"spread_pct":3.0636004571899127},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":4.432046785950661,"wave_count":2},"fallback_work":721723232,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":453.5256652832031,"n":8,"spread_pct":4.194109343913238},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.216818138957024,"wave_count":1},"fallback_work":604479328,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":430.04345703125,"n":8,"spread_pct":2.02020144675604},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":9.784698486328125,"wave_count":3},"fallback_work":838967136,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":458.2540740966797,"n":8,"spread_pct":3.20027820862177}],"cell":"cal-hetero|Qwen/Qwen3-1.7B|L28|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19380,"summary_ms":0.0004470348358154297,"wave_count":0},"fallback_work":560404992,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":445.8825988769531,"n":8,"spread_pct":1.2347542185063105},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":722147840,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":544.3458251953125,"n":8,"spread_pct":1.9700485347301504},{"cp_plan":{"max_rank_tokens":20916,"summary_ms":0.00024028122425079346,"wave_count":0},"fallback_work":604412416,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":465.85182189941406,"n":8,"spread_pct":1.378418290298432},{"cp_plan":{"max_rank_tokens":29108,"summary_ms":0.00039674341678619385,"wave_count":0},"fallback_work":839883264,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":629.9834289550781,"n":8,"spread_pct":0.606860425444637}],"cell":"cal-hetero|Qwen/Qwen3-1.7B|L28|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":28.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":6144,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":2048,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":8},"hidden_size":2048,"model":"Qwen/Qwen3-1.7B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}} + ], + "fit_arguments": { + "exclude_cells": ["cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g7","cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-hetero2|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-hetero3|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-hetero|Qwen/Qwen3-1.7B|L28|tp2|cp2"], + "holdout": "", + "incumbent": "depth_one", + "objective": "regret", + "require_complete": 8, + "rerank_shapes": ["tp1cp4"], + "shortlist_size": 3, + "terms": ["token_per_rank","token_cp_exchange","token_tp_collective","gdn_token_per_rank","attention_token_cp_exchange","tiny_segment_per_layer","level_cp_per_layer","level_tp_per_layer","gdn_level","gdn_level_tp"] + }, + "integer_table_milli_us": {"attention_token_cp_exchange":0,"gdn_level":0,"gdn_level_tp":0,"gdn_token_per_rank":0,"level_cp_per_layer":0,"level_tp_per_layer":0,"tiny_segment_per_layer":0,"token_cp_exchange":46,"token_per_rank":259,"token_tp_collective":573}, + "integer_table_sha256": "fcc224b1a9df7b5d7c66b9bb688c0901963342003429699d42d6d401a6ae2c5b", + "manifest": { + "excluded": ["cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g7","cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-hetero2|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-hetero3|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-hetero|Qwen/Qwen3-1.7B|L28|tp2|cp2"], + "expected": ["cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g0","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g1","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g2","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g3","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g4","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g5","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g6","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g7","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g7","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g0","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g1","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g2","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g3","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g4","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g5","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g6","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g7","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g0","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g1","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g2","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g3","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g4","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g5","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g6","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g7","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g7","cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp1|cp1","cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp1|cp2","cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp1|cp4","cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp2|cp1","cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp1|cp1","cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp1|cp2","cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp1|cp4","cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp2|cp1","cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp1|cp1","cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp1|cp2","cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp1|cp4","cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp2|cp1","cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-hetero2|Qwen/Qwen3-1.7B|L28|tp1|cp1","cal-hetero2|Qwen/Qwen3-1.7B|L28|tp1|cp2","cal-hetero2|Qwen/Qwen3-1.7B|L28|tp1|cp4","cal-hetero2|Qwen/Qwen3-1.7B|L28|tp2|cp1","cal-hetero2|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-hetero3|Qwen/Qwen3-1.7B|L28|tp1|cp1","cal-hetero3|Qwen/Qwen3-1.7B|L28|tp1|cp2","cal-hetero3|Qwen/Qwen3-1.7B|L28|tp1|cp4","cal-hetero3|Qwen/Qwen3-1.7B|L28|tp2|cp1","cal-hetero3|Qwen/Qwen3-1.7B|L28|tp2|cp2","cal-hetero|Qwen/Qwen3-1.7B|L28|tp1|cp1","cal-hetero|Qwen/Qwen3-1.7B|L28|tp1|cp2","cal-hetero|Qwen/Qwen3-1.7B|L28|tp1|cp4","cal-hetero|Qwen/Qwen3-1.7B|L28|tp2|cp1","cal-hetero|Qwen/Qwen3-1.7B|L28|tp2|cp2"], + "path": "trainer_rank_cost_calibration_manifest_dense-attn-h2048-h200-bf16.json" + }, + "measured_envelope": {"device_names":["NVIDIA H200"],"hidden_sizes":[2048],"param_dtypes":["torch.bfloat16"]}, + "metrics": { + "integer_all": {"cells":42,"clear_misses":[],"max_regret_pct":0.9775346211851169,"median_regret_pct":0.0,"ordered_pairs":2792,"p95_regret_pct":0.723341595088529,"pairwise_accuracy":0.964541547277937}, + "test": {"cells":12,"clear_misses":[],"max_regret_pct":0.9775346211851169,"median_regret_pct":0.0,"ordered_pairs":596,"p95_regret_pct":0.9775346211851169,"pairwise_accuracy":0.9983221476510067}, + "train": {"cells":30,"clear_misses":[],"max_regret_pct":0.692725506764514,"median_regret_pct":0.0,"ordered_pairs":2196,"p95_regret_pct":0.6340678453401267,"pairwise_accuracy":0.9553734061930783} + }, + "reranker": {"incumbent":"depth_one","max_rank_token_per_layer_milli_us":678,"metrics":{"all":{"cells":14,"clear_misses":[],"max_regret_pct":2.6274480101219906,"median_regret_pct":0.0,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":2.878283716142653,"savings_vs_fallback_mean_pct":1.9714841301057868,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":26,"p95_regret_pct":2.6274480101219906,"pairwise_accuracy":0.9615384615384616,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g0":{"best":"minimum_effective_span_46","best_lead_pct":0.12127059798202275,"cheap_regret_pct":0.1480474663655057,"fallback_regret_pct":0.12127059798202275,"irrecoverable_pct":0.0,"planning_ms":8.176727220416069,"regret_pct":0.1480474663655057,"selected":"depth_one","shortlist":["depth_one","uniform_depth_2","minimum_effective_span_46"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g1":{"best":"depth_one","best_lead_pct":11.622434338923622,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.602181449532509,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g2":{"best":"no_sharing","best_lead_pct":0.0,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":0.0,"regret_pct":0.0,"selected":"no_sharing","shortlist":["no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g3":{"best":"depth_one","best_lead_pct":5.596434642047594,"cheap_regret_pct":10.071593513949747,"fallback_regret_pct":8.2946735594026,"irrecoverable_pct":0.0,"planning_ms":22.739894688129425,"regret_pct":0.0,"selected":"depth_one","shortlist":["minimum_effective_span_278","uniform_depth_3","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g4":{"best":"uniform_depth_2","best_lead_pct":6.5084130827810505,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.731650844216347,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_6717"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g5":{"best":"uniform_depth_4","best_lead_pct":0.005097524232328771,"cheap_regret_pct":26.390080792822364,"fallback_regret_pct":0.005097524232328771,"irrecoverable_pct":0.26845003144915874,"planning_ms":10.200833901762962,"regret_pct":0.6116534473653729,"selected":"minimum_effective_span_22","shortlist":["depth_one","minimum_effective_span_22","minimum_effective_span_179"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g6":{"best":"uniform_depth_10","best_lead_pct":0.09172247467889252,"cheap_regret_pct":2.6274480101219906,"fallback_regret_pct":16.607937984623444,"irrecoverable_pct":2.6274480101219906,"planning_ms":11.008212342858315,"regret_pct":2.6274480101219906,"selected":"minimum_effective_span_197","shortlist":["minimum_effective_span_197","minimum_effective_span_133","minimum_effective_span_300","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g7":{"best":"minimum_effective_span_134","best_lead_pct":0.08079813873769395,"cheap_regret_pct":0.7092883387512732,"fallback_regret_pct":2.2222842512543584,"irrecoverable_pct":0.6538755148605172,"planning_ms":9.203648194670677,"regret_pct":0.7092883387512732,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_87","minimum_effective_span_53"]},"cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":14.346447850733252,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.280426919460297,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":4.057563172369688,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":11.940130963921547,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":17.199003953850024,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":9.619837626814842,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp1|cp4":{"best":"depth_one","best_lead_pct":0.4153611798035509,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":25.870174169540405,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_257","minimum_effective_span_1025"]},"cal-hetero3|Qwen/Qwen3-1.7B|L28|tp1|cp4":{"best":"depth_one","best_lead_pct":1.1508697451509442,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":19.830884411931038,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_769","no_sharing"]},"cal-hetero|Qwen/Qwen3-1.7B|L28|tp1|cp4":{"best":"minimum_effective_span_513","best_lead_pct":4.445951166590404,"cheap_regret_pct":4.445951166590404,"fallback_regret_pct":4.445951166590404,"irrecoverable_pct":0.0,"planning_ms":8.648864924907684,"regret_pct":0.0,"selected":"minimum_effective_span_513","shortlist":["depth_one","minimum_effective_span_513","minimum_effective_span_2049"]}},"planning_cost":{"max_cell":"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp1|cp4","max_ms":25.870174169540405,"max_pct_of_cell":4.757694195368482,"mean_pct_of_cell":2.632820159302704,"median_pct_of_cell":2.6227973418960553},"recall":{"best_in_shortlist":11,"cells":14,"clear_winner_cells":6,"clear_winners_in_shortlist":6,"max_irrecoverable_pct":2.6274480101219906}},"by_shape":{"tp1cp4":{"cells":14,"clear_misses":[],"max_regret_pct":2.6274480101219906,"median_regret_pct":0.0,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":2.878283716142653,"savings_vs_fallback_mean_pct":1.9714841301057868,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":26,"p95_regret_pct":2.6274480101219906,"pairwise_accuracy":0.9615384615384616,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g0":{"best":"minimum_effective_span_46","best_lead_pct":0.12127059798202275,"cheap_regret_pct":0.1480474663655057,"fallback_regret_pct":0.12127059798202275,"irrecoverable_pct":0.0,"planning_ms":8.176727220416069,"regret_pct":0.1480474663655057,"selected":"depth_one","shortlist":["depth_one","uniform_depth_2","minimum_effective_span_46"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g1":{"best":"depth_one","best_lead_pct":11.622434338923622,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.602181449532509,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g2":{"best":"no_sharing","best_lead_pct":0.0,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":0.0,"regret_pct":0.0,"selected":"no_sharing","shortlist":["no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g3":{"best":"depth_one","best_lead_pct":5.596434642047594,"cheap_regret_pct":10.071593513949747,"fallback_regret_pct":8.2946735594026,"irrecoverable_pct":0.0,"planning_ms":22.739894688129425,"regret_pct":0.0,"selected":"depth_one","shortlist":["minimum_effective_span_278","uniform_depth_3","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g4":{"best":"uniform_depth_2","best_lead_pct":6.5084130827810505,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.731650844216347,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_6717"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g5":{"best":"uniform_depth_4","best_lead_pct":0.005097524232328771,"cheap_regret_pct":26.390080792822364,"fallback_regret_pct":0.005097524232328771,"irrecoverable_pct":0.26845003144915874,"planning_ms":10.200833901762962,"regret_pct":0.6116534473653729,"selected":"minimum_effective_span_22","shortlist":["depth_one","minimum_effective_span_22","minimum_effective_span_179"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g6":{"best":"uniform_depth_10","best_lead_pct":0.09172247467889252,"cheap_regret_pct":2.6274480101219906,"fallback_regret_pct":16.607937984623444,"irrecoverable_pct":2.6274480101219906,"planning_ms":11.008212342858315,"regret_pct":2.6274480101219906,"selected":"minimum_effective_span_197","shortlist":["minimum_effective_span_197","minimum_effective_span_133","minimum_effective_span_300","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g7":{"best":"minimum_effective_span_134","best_lead_pct":0.08079813873769395,"cheap_regret_pct":0.7092883387512732,"fallback_regret_pct":2.2222842512543584,"irrecoverable_pct":0.6538755148605172,"planning_ms":9.203648194670677,"regret_pct":0.7092883387512732,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_87","minimum_effective_span_53"]},"cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":14.346447850733252,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.280426919460297,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":4.057563172369688,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":11.940130963921547,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":17.199003953850024,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":9.619837626814842,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp1|cp4":{"best":"depth_one","best_lead_pct":0.4153611798035509,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":25.870174169540405,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_257","minimum_effective_span_1025"]},"cal-hetero3|Qwen/Qwen3-1.7B|L28|tp1|cp4":{"best":"depth_one","best_lead_pct":1.1508697451509442,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":19.830884411931038,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_769","no_sharing"]},"cal-hetero|Qwen/Qwen3-1.7B|L28|tp1|cp4":{"best":"minimum_effective_span_513","best_lead_pct":4.445951166590404,"cheap_regret_pct":4.445951166590404,"fallback_regret_pct":4.445951166590404,"irrecoverable_pct":0.0,"planning_ms":8.648864924907684,"regret_pct":0.0,"selected":"minimum_effective_span_513","shortlist":["depth_one","minimum_effective_span_513","minimum_effective_span_2049"]}},"planning_cost":{"max_cell":"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp1|cp4","max_ms":25.870174169540405,"max_pct_of_cell":4.757694195368482,"mean_pct_of_cell":2.632820159302704,"median_pct_of_cell":2.6227973418960553},"recall":{"best_in_shortlist":11,"cells":14,"clear_winner_cells":6,"clear_winners_in_shortlist":6,"max_irrecoverable_pct":2.6274480101219906}}},"held_out":{"cells":4,"clear_misses":[],"max_regret_pct":0.7092883387512732,"median_regret_pct":0.30582672368268643,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":8.962505214851685,"savings_vs_fallback_mean_pct":2.3002783871931602,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":8,"p95_regret_pct":0.7092883387512732,"pairwise_accuracy":1.0,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g1":{"best":"depth_one","best_lead_pct":11.622434338923622,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.602181449532509,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g3":{"best":"depth_one","best_lead_pct":5.596434642047594,"cheap_regret_pct":10.071593513949747,"fallback_regret_pct":8.2946735594026,"irrecoverable_pct":0.0,"planning_ms":22.739894688129425,"regret_pct":0.0,"selected":"depth_one","shortlist":["minimum_effective_span_278","uniform_depth_3","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g5":{"best":"uniform_depth_4","best_lead_pct":0.005097524232328771,"cheap_regret_pct":26.390080792822364,"fallback_regret_pct":0.005097524232328771,"irrecoverable_pct":0.26845003144915874,"planning_ms":10.200833901762962,"regret_pct":0.6116534473653729,"selected":"minimum_effective_span_22","shortlist":["depth_one","minimum_effective_span_22","minimum_effective_span_179"]},"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g7":{"best":"minimum_effective_span_134","best_lead_pct":0.08079813873769395,"cheap_regret_pct":0.7092883387512732,"fallback_regret_pct":2.2222842512543584,"irrecoverable_pct":0.6538755148605172,"planning_ms":9.203648194670677,"regret_pct":0.7092883387512732,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_87","minimum_effective_span_53"]}},"planning_cost":{"max_cell":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g3","max_ms":22.739894688129425,"max_pct_of_cell":4.757519292742499,"mean_pct_of_cell":2.6129229349627145,"median_pct_of_cell":2.066589650214124},"recall":{"best_in_shortlist":2,"cells":4,"clear_winner_cells":1,"clear_winners_in_shortlist":1,"max_irrecoverable_pct":0.6538755148605172}}},"shapes":["tp1cp4"],"shortlist_size":3,"shortlist_table_milli_us":{"attention_token_cp_exchange":3,"gdn_level":0,"gdn_level_tp":0,"gdn_token_per_rank":0,"level_cp_per_layer":69746,"level_tp_per_layer":0,"tiny_segment_per_layer":0,"token_cp_exchange":57,"token_per_rank":1656,"token_tp_collective":143},"wave_per_layer_milli_us":1144207}, + "schema": "art.dev.trainer_rank_cost_calibration_certificate.v1", + "table_id": "dense-attn-h2048-h200-bf16" +} diff --git a/dev/trainer_rank_cost_calibration_certificate_dense-attn-h4096-h200-bf16.json b/dev/trainer_rank_cost_calibration_certificate_dense-attn-h4096-h200-bf16.json new file mode 100644 index 000000000..73bbbd119 --- /dev/null +++ b/dev/trainer_rank_cost_calibration_certificate_dense-attn-h4096-h200-bf16.json @@ -0,0 +1,87 @@ +{ + "admitted": {"device_classes":[{"capability":[9,0],"memory_class":"hbm-141g"}],"geometries":[{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8}],"param_dtypes":["torch.bfloat16"],"shapes":[[1,1,1,1],[1,2,1,1],[1,4,1,1],[2,1,1,1]]}, + "cells": [ + {"candidates":[{"cp_plan":{"max_rank_tokens":12204,"summary_ms":0.0005252659320831299,"wave_count":0},"fallback_work":452204032,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1521.0401611328125,"n":8,"spread_pct":0.35184373871671937},{"cp_plan":{"max_rank_tokens":23926,"summary_ms":0.0002868473529815674,"wave_count":0},"fallback_work":885824256,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":2970.0048828125,"n":8,"spread_pct":0.2971523550078339},{"cp_plan":{"max_rank_tokens":35778,"summary_ms":0.0004246830940246582,"wave_count":0},"fallback_work":1324253440,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":4375.22900390625,"n":8,"spread_pct":0.24177305897715842},{"cp_plan":{"max_rank_tokens":12154,"summary_ms":0.00036135315895080566,"wave_count":0},"fallback_work":450583808,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1520.5737915039062,"n":8,"spread_pct":0.6153554431906676},{"cp_plan":{"max_rank_tokens":47640,"summary_ms":0.0012721866369247437,"wave_count":0},"fallback_work":1762823168,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5848.46484375,"n":8,"spread_pct":0.310628390062296},{"cp_plan":{"max_rank_tokens":12114,"summary_ms":0.0005159527063369751,"wave_count":0},"fallback_work":449104128,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1516.89453125,"n":8,"spread_pct":0.6826594991308826},{"cp_plan":{"max_rank_tokens":12109,"summary_ms":0.00034086406230926514,"wave_count":0},"fallback_work":449148544,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1517.43603515625,"n":8,"spread_pct":1.1010924768052608}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11920,"summary_ms":0.00039301812648773193,"wave_count":0},"fallback_work":441436160,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1467.8790893554688,"n":8,"spread_pct":0.4564375377005257},{"cp_plan":{"max_rank_tokens":23728,"summary_ms":0.0009872019290924072,"wave_count":0},"fallback_work":878008320,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2895.2342529296875,"n":8,"spread_pct":0.5894907693429214}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12138,"summary_ms":0.0006780028343200684,"wave_count":0},"fallback_work":449139968,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1518.8861694335938,"n":8,"spread_pct":0.2957553759064892}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":20701,"summary_ms":0.00040605664253234863,"wave_count":0},"fallback_work":766656128,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":2615.974609375,"n":8,"spread_pct":0.7146315568632926},{"cp_plan":{"max_rank_tokens":16154,"summary_ms":0.00023655593395233154,"wave_count":0},"fallback_work":598682880,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":2021.560302734375,"n":8,"spread_pct":0.7659374146695698},{"cp_plan":{"max_rank_tokens":14771,"summary_ms":0.0002514570951461792,"wave_count":0},"fallback_work":547752320,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1857.0800170898438,"n":8,"spread_pct":0.9653064761294954},{"cp_plan":{"max_rank_tokens":21963,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":813569408,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":2641.6346435546875,"n":8,"spread_pct":0.9009776263892443},{"cp_plan":{"max_rank_tokens":27497,"summary_ms":0.0001862645149230957,"wave_count":0},"fallback_work":1018283136,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":3229.3441162109375,"n":8,"spread_pct":1.0386551212829647},{"cp_plan":{"max_rank_tokens":35243,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":1304593792,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":4097.460693359375,"n":8,"spread_pct":0.7016004517374786},{"cp_plan":{"max_rank_tokens":46153,"summary_ms":0.0008959323167800903,"wave_count":0},"fallback_work":1707947136,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":5492.322509765625,"n":8,"spread_pct":0.7742341682301284},{"cp_plan":{"max_rank_tokens":15600,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":578418688,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1934.942138671875,"n":8,"spread_pct":0.9379948805902146},{"cp_plan":{"max_rank_tokens":14494,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":537734912,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1817.9356079101562,"n":8,"spread_pct":1.0445304838597422}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16800,"summary_ms":0.00038743019104003906,"wave_count":0},"fallback_work":622088192,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":2114.9541015625,"n":8,"spread_pct":0.607421204933433},{"cp_plan":{"max_rank_tokens":18971,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":702397824,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":2275.2515869140625,"n":8,"spread_pct":0.9144979406198693},{"cp_plan":{"max_rank_tokens":30232,"summary_ms":0.000998377799987793,"wave_count":0},"fallback_work":1118735360,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":3628.3172607421875,"n":8,"spread_pct":0.6743290378635526},{"cp_plan":{"max_rank_tokens":12255,"summary_ms":0.00024400651454925537,"wave_count":0},"fallback_work":454188928,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":1523.4889526367188,"n":8,"spread_pct":1.0369120946756487}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13078,"summary_ms":0.00046938657760620117,"wave_count":0},"fallback_work":484666112,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1650.3259887695312,"n":8,"spread_pct":0.42606698243100566},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":926128640,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":3095.813720703125,"n":8,"spread_pct":0.2995870888839697},{"cp_plan":{"max_rank_tokens":36716,"summary_ms":0.00028312206268310547,"wave_count":0},"fallback_work":1359083008,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":4457.3603515625,"n":8,"spread_pct":0.11359809759884075},{"cp_plan":{"max_rank_tokens":48890,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":1809423616,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":6021.281005859375,"n":8,"spread_pct":0.27246298093022686},{"cp_plan":{"max_rank_tokens":12899,"summary_ms":0.0002905726432800293,"wave_count":0},"fallback_work":478273920,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1635.1917724609375,"n":8,"spread_pct":0.43318307630530306},{"cp_plan":{"max_rank_tokens":12722,"summary_ms":0.00026263296604156494,"wave_count":0},"fallback_work":471726336,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1615.1865844726562,"n":8,"spread_pct":0.8834677960733194},{"cp_plan":{"max_rank_tokens":61066,"summary_ms":0.0009313225746154785,"wave_count":0},"fallback_work":2259608832,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":7635.11181640625,"n":8,"spread_pct":0.15680409477467414},{"cp_plan":{"max_rank_tokens":13015,"summary_ms":0.00028870999813079834,"wave_count":0},"fallback_work":482564992,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1655.3746948242188,"n":8,"spread_pct":0.44183875330045264},{"cp_plan":{"max_rank_tokens":12701,"summary_ms":0.00032782554626464844,"wave_count":0},"fallback_work":471178880,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1614.0620727539062,"n":8,"spread_pct":0.35408104725947376},{"cp_plan":{"max_rank_tokens":12700,"summary_ms":0.0002868473529815674,"wave_count":0},"fallback_work":471371264,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1602.7064208984375,"n":8,"spread_pct":0.43890145232395367}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9058,"summary_ms":0.0005904585123062134,"wave_count":0},"fallback_work":337006848,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1077.3926391601562,"n":8,"spread_pct":0.7541014755211038},{"cp_plan":{"max_rank_tokens":6403,"summary_ms":0.0003110617399215698,"wave_count":0},"fallback_work":239710592,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":779.2690734863281,"n":8,"spread_pct":1.4891508700048},{"cp_plan":{"max_rank_tokens":6443,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":241190272,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":781.2079772949219,"n":8,"spread_pct":1.1657736433713641},{"cp_plan":{"max_rank_tokens":6668,"summary_ms":0.0002551823854446411,"wave_count":0},"fallback_work":249284096,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":798.2217407226562,"n":8,"spread_pct":2.6209687748149286},{"cp_plan":{"max_rank_tokens":6755,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":252502400,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":819.9093933105469,"n":8,"spread_pct":0.8796956808127923},{"cp_plan":{"max_rank_tokens":6799,"summary_ms":0.000260770320892334,"wave_count":0},"fallback_work":254130048,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":823.4574890136719,"n":8,"spread_pct":0.8821462210264924},{"cp_plan":{"max_rank_tokens":7040,"summary_ms":0.00025890767574310303,"wave_count":0},"fallback_work":262815744,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":838.5472717285156,"n":8,"spread_pct":0.8212386876150418},{"cp_plan":{"max_rank_tokens":7291,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":272100736,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":871.9953002929688,"n":8,"spread_pct":0.5633463292720238},{"cp_plan":{"max_rank_tokens":7367,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":274912128,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":882.2160339355469,"n":8,"spread_pct":1.6166481397916284},{"cp_plan":{"max_rank_tokens":7666,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":285743360,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":912.9239807128906,"n":8,"spread_pct":0.6773460787312776},{"cp_plan":{"max_rank_tokens":5988,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":225735168,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":718.9390869140625,"n":8,"spread_pct":1.3855560077690159},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":288406784,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":926.9967651367188,"n":8,"spread_pct":1.88227284653621},{"cp_plan":{"max_rank_tokens":6024,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":226837504,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":716.4728088378906,"n":8,"spread_pct":1.0366658870510301},{"cp_plan":{"max_rank_tokens":6102,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":229264128,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":742.0941162109375,"n":8,"spread_pct":1.3229707356270979},{"cp_plan":{"max_rank_tokens":8016,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":298690560,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":956.4167175292969,"n":8,"spread_pct":3.8001442444084708},{"cp_plan":{"max_rank_tokens":13361,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":496183424,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":1537.9149169921875,"n":8,"spread_pct":34.62769947674341},{"cp_plan":{"max_rank_tokens":17592,"summary_ms":0.00025890767574310303,"wave_count":0},"fallback_work":652696576,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":1942.7374267578125,"n":8,"spread_pct":0.8120249539661092},{"cp_plan":{"max_rank_tokens":22066,"summary_ms":0.00025890767574310303,"wave_count":0},"fallback_work":818198784,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":2420.5699462890625,"n":8,"spread_pct":0.8637315601869939},{"cp_plan":{"max_rank_tokens":26682,"summary_ms":0.00025704503059387207,"wave_count":0},"fallback_work":988953856,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":2952.7677001953125,"n":8,"spread_pct":0.5009617359044722},{"cp_plan":{"max_rank_tokens":31360,"summary_ms":0.0002942979335784912,"wave_count":0},"fallback_work":1162002432,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":3409.3206787109375,"n":8,"spread_pct":1.0762158754892512},{"cp_plan":{"max_rank_tokens":35986,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":1333127424,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":3898.985107421875,"n":8,"spread_pct":0.7389618304492924},{"cp_plan":{"max_rank_tokens":40782,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":1510541056,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":4403.422119140625,"n":8,"spread_pct":0.5388986071901087},{"cp_plan":{"max_rank_tokens":45522,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":1685883136,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":4898.90283203125,"n":8,"spread_pct":0.3840145568308752},{"cp_plan":{"max_rank_tokens":50326,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":1863592704,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":5418.790771484375,"n":8,"spread_pct":0.8901857761577},{"cp_plan":{"max_rank_tokens":55336,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":2048922624,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":5848.02734375,"n":8,"spread_pct":0.6892264685941775},{"cp_plan":{"max_rank_tokens":8411,"summary_ms":0.0003110617399215698,"wave_count":0},"fallback_work":313302400,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":1006.4861755371094,"n":8,"spread_pct":5.252533926143536},{"cp_plan":{"max_rank_tokens":60463,"summary_ms":0.00029616057872772217,"wave_count":0},"fallback_work":2238580608,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":6473.379150390625,"n":8,"spread_pct":0.6832897095341758},{"cp_plan":{"max_rank_tokens":6223,"summary_ms":0.00027008354663848877,"wave_count":0},"fallback_work":233510784,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":759.6538696289062,"n":8,"spread_pct":1.1154114973049138},{"cp_plan":{"max_rank_tokens":6304,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":236277760,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":758.3622741699219,"n":8,"spread_pct":0.6826394269782131},{"cp_plan":{"max_rank_tokens":6322,"summary_ms":0.0003520399332046509,"wave_count":0},"fallback_work":236943616,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":764.032958984375,"n":8,"spread_pct":0.7666851255432613},{"cp_plan":{"max_rank_tokens":6333,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":237350528,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":766.0141296386719,"n":8,"spread_pct":0.8631136175965411},{"cp_plan":{"max_rank_tokens":6310,"summary_ms":0.00027008354663848877,"wave_count":0},"fallback_work":236499712,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":765.8975219726562,"n":8,"spread_pct":1.6430376145608248},{"cp_plan":{"max_rank_tokens":65842,"summary_ms":0.0011194497346878052,"wave_count":0},"fallback_work":2437331200,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":7175.21923828125,"n":8,"spread_pct":0.5504037503209458},{"cp_plan":{"max_rank_tokens":6245,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":235012736,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":760.82861328125,"n":8,"spread_pct":1.6879340408695547},{"cp_plan":{"max_rank_tokens":6083,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":229249408,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":739.7401428222656,"n":8,"spread_pct":2.9070778415915473},{"cp_plan":{"max_rank_tokens":5957,"summary_ms":0.00030919909477233887,"wave_count":0},"fallback_work":224817792,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":723.4076232910156,"n":8,"spread_pct":1.287149876810939},{"cp_plan":{"max_rank_tokens":8629,"summary_ms":0.00032223761081695557,"wave_count":0},"fallback_work":321366656,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1036.0068969726562,"n":8,"spread_pct":0.6020877772787775},{"cp_plan":{"max_rank_tokens":8089,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":301620352,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":966.6283264160156,"n":8,"spread_pct":1.1152765048022735},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":288865536,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":929.8396911621094,"n":8,"spread_pct":1.1413704039454937},{"cp_plan":{"max_rank_tokens":7490,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":279920896,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":902.3811340332031,"n":8,"spread_pct":1.417101733305489},{"cp_plan":{"max_rank_tokens":7273,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":272123008,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":868.6482849121094,"n":8,"spread_pct":1.1936045091835203},{"cp_plan":{"max_rank_tokens":6973,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":261254784,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":838.66796875,"n":8,"spread_pct":0.8790722709467673},{"cp_plan":{"max_rank_tokens":6793,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":254825600,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":824.8542785644531,"n":8,"spread_pct":0.29003110568069457},{"cp_plan":{"max_rank_tokens":6521,"summary_ms":0.0002868473529815674,"wave_count":0},"fallback_work":244993152,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":785.4738464355469,"n":8,"spread_pct":0.8815852511773553}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12291,"summary_ms":0.0004954636096954346,"wave_count":0},"fallback_work":455946624,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1570.7865600585938,"n":8,"spread_pct":1.3354799178901489},{"cp_plan":{"max_rank_tokens":23563,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":872920448,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":2898.6610107421875,"n":8,"spread_pct":0.708385400367753},{"cp_plan":{"max_rank_tokens":34902,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":1292372736,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":4269.9921875,"n":8,"spread_pct":0.8787706359240265},{"cp_plan":{"max_rank_tokens":46161,"summary_ms":0.0005718320608139038,"wave_count":0},"fallback_work":1708865664,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":5601.077880859375,"n":8,"spread_pct":0.8766710334728662},{"cp_plan":{"max_rank_tokens":57385,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":2124063872,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":6949.9541015625,"n":8,"spread_pct":0.6657389964215422},{"cp_plan":{"max_rank_tokens":68715,"summary_ms":0.00027008354663848877,"wave_count":0},"fallback_work":2543183232,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":8370.322265625,"n":8,"spread_pct":2.961425516350667},{"cp_plan":{"max_rank_tokens":80181,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":2967333504,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":9710.4384765625,"n":8,"spread_pct":0.4464931775959123},{"cp_plan":{"max_rank_tokens":11890,"summary_ms":0.00029616057872772217,"wave_count":0},"fallback_work":442030336,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":1524.383056640625,"n":8,"spread_pct":16.994452001185802},{"cp_plan":{"max_rank_tokens":12153,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":451071104,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1546.9610595703125,"n":8,"spread_pct":0.8917985544522657},{"cp_plan":{"max_rank_tokens":11903,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":442281856,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1526.7876586914062,"n":8,"spread_pct":0.9292785588614844},{"cp_plan":{"max_rank_tokens":11876,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":441741824,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":1516.7489013671875,"n":8,"spread_pct":1.4517102210047688},{"cp_plan":{"max_rank_tokens":11939,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":443384192,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1527.3109741210938,"n":8,"spread_pct":0.8082014867406642},{"cp_plan":{"max_rank_tokens":11981,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":444937856,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1527.7434692382812,"n":8,"spread_pct":0.9520845907135317},{"cp_plan":{"max_rank_tokens":12033,"summary_ms":0.00025704503059387207,"wave_count":0},"fallback_work":446632064,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1540.3735961914062,"n":8,"spread_pct":1.8748860078388965},{"cp_plan":{"max_rank_tokens":11881,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":441697408,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":1523.6053466796875,"n":8,"spread_pct":20.941266487065725},{"cp_plan":{"max_rank_tokens":12025,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":446336128,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1536.6716918945312,"n":8,"spread_pct":0.7836888434202735},{"cp_plan":{"max_rank_tokens":91657,"summary_ms":0.0010337680578231812,"wave_count":0},"fallback_work":3391624320,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":11351.02978515625,"n":8,"spread_pct":0.682301462099757},{"cp_plan":{"max_rank_tokens":12225,"summary_ms":0.000303611159324646,"wave_count":0},"fallback_work":453734528,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1562.5746459960938,"n":8,"spread_pct":0.8118362157932846},{"cp_plan":{"max_rank_tokens":12215,"summary_ms":0.00031851232051849365,"wave_count":0},"fallback_work":453593984,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1564.043701171875,"n":8,"spread_pct":0.9327113228602428},{"cp_plan":{"max_rank_tokens":12123,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":450420096,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1545.0782470703125,"n":8,"spread_pct":0.6094594881791753},{"cp_plan":{"max_rank_tokens":11973,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":445100672,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":1527.4796752929688,"n":8,"spread_pct":0.8796531734487985},{"cp_plan":{"max_rank_tokens":11879,"summary_ms":0.00027008354663848877,"wave_count":0},"fallback_work":441852800,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":1522.3322143554688,"n":8,"spread_pct":1.0002609665802806},{"cp_plan":{"max_rank_tokens":11874,"summary_ms":0.0002868473529815674,"wave_count":0},"fallback_work":441897216,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":1520.6328735351562,"n":8,"spread_pct":0.9788131989304508}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.062079221010208,"wave_count":2},"fallback_work":451814848,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1320.65625,"n":8,"spread_pct":0.9071695091218854},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":12.871799990534782,"wave_count":1},"fallback_work":884872416,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":2020.9584350585938,"n":8,"spread_pct":0.38924378362570966},{"cp_plan":{"max_rank_tokens":20480,"summary_ms":5.354784429073334,"wave_count":1},"fallback_work":1322732704,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":2690.69580078125,"n":8,"spread_pct":0.6967278191177098},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.58001734316349,"wave_count":2},"fallback_work":450262560,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1318.2363891601562,"n":8,"spread_pct":1.2585248114581058},{"cp_plan":{"max_rank_tokens":27136,"summary_ms":8.789842948317528,"wave_count":1},"fallback_work":1760667520,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3590.1759033203125,"n":8,"spread_pct":0.7879915626685928},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.727043375372887,"wave_count":2},"fallback_work":448784800,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1318.6725463867188,"n":8,"spread_pct":0.8222306080902726},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.645686760544777,"wave_count":2},"fallback_work":448895072,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1317.5783081054688,"n":8,"spread_pct":0.7854838813532259}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.329864263534546,"wave_count":2},"fallback_work":440995072,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1303.048828125,"n":8,"spread_pct":0.8645495615001476},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":8.714893832802773,"wave_count":1},"fallback_work":876934912,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2010.9064331054688,"n":8,"spread_pct":0.4464909693055366}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.279917269945145,"wave_count":2},"fallback_work":448590112,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1272.412353515625,"n":8,"spread_pct":0.6130702227502547}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13824,"summary_ms":8.728539571166039,"wave_count":1},"fallback_work":765891936,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":1903.1765747070312,"n":8,"spread_pct":0.9872089814376375},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.340999156236649,"wave_count":2},"fallback_work":598202400,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":1540.1211547851562,"n":8,"spread_pct":0.8032228782985371},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":5.9320032596588135,"wave_count":2},"fallback_work":547403840,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1487.177490234375,"n":8,"spread_pct":0.4763293118140931},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.221052259206772,"wave_count":1},"fallback_work":812810176,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":1522.93994140625,"n":8,"spread_pct":18.747629434100432},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":7.859673351049423,"wave_count":4},"fallback_work":1017258272,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":1858.4828491210938,"n":8,"spread_pct":0.35548831974690587},{"cp_plan":{"max_rank_tokens":17835,"summary_ms":5.492573603987694,"wave_count":1},"fallback_work":1303131584,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":2327.9813232421875,"n":8,"spread_pct":0.3512066574158861},{"cp_plan":{"max_rank_tokens":24649,"summary_ms":6.297716870903969,"wave_count":1},"fallback_work":1705895712,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":3170.37646484375,"n":8,"spread_pct":0.4658527866690713},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":7.233625277876854,"wave_count":3},"fallback_work":578030336,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1491.16650390625,"n":8,"spread_pct":0.7926153005868871},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":7.305623963475227,"wave_count":3},"fallback_work":537465184,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1383.1282958984375,"n":8,"spread_pct":1.1123783789842487}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11776,"summary_ms":6.263164803385735,"wave_count":2},"fallback_work":621445632,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":1611.1788330078125,"n":8,"spread_pct":0.6512573096734644},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":6.198117509484291,"wave_count":1},"fallback_work":701651136,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":1789.2288818359375,"n":8,"spread_pct":0.47526284229980525},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":7.548833265900612,"wave_count":1},"fallback_work":1117382528,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":2303.81982421875,"n":8,"spread_pct":0.35473100783918016},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.784271284937859,"wave_count":2},"fallback_work":453830144,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":1308.7527465820312,"n":8,"spread_pct":0.5492802763470204}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.681196719408035,"wave_count":2},"fallback_work":484267744,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1389.1905517578125,"n":8,"spread_pct":0.9693900370721239},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":11.687040328979492,"wave_count":1},"fallback_work":925157440,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":1798.4957885742188,"n":8,"spread_pct":0.7212108778919015},{"cp_plan":{"max_rank_tokens":21504,"summary_ms":8.754264563322067,"wave_count":1},"fallback_work":1357550016,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":2885.9970703125,"n":8,"spread_pct":0.36030658544203037},{"cp_plan":{"max_rank_tokens":25850,"summary_ms":16.055410727858543,"wave_count":1},"fallback_work":1807306272,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":3564.276611328125,"n":8,"spread_pct":0.4007595718609901},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.586470037698746,"wave_count":2},"fallback_work":477949760,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1385.1342163085938,"n":8,"spread_pct":0.9044665870187322},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.329208940267563,"wave_count":2},"fallback_work":471410592,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1385.2432250976562,"n":8,"spread_pct":0.699820341406404},{"cp_plan":{"max_rank_tokens":33280,"summary_ms":7.605815306305885,"wave_count":1},"fallback_work":2256841504,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":4453.968017578125,"n":8,"spread_pct":0.43025846039124277},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.708691224455833,"wave_count":2},"fallback_work":482235264,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1386.4559326171875,"n":8,"spread_pct":47.124468220077276},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.429301902651787,"wave_count":2},"fallback_work":470929760,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1383.7429809570312,"n":8,"spread_pct":0.918679146204807},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.56340117752552,"wave_count":2},"fallback_work":471187648,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1387.4168090820312,"n":8,"spread_pct":0.6082325547654555}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7680,"summary_ms":6.632447242736816,"wave_count":1},"fallback_work":337063584,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1026.3575439453125,"n":8,"spread_pct":1.5609506797221335},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.045065656304359,"wave_count":2},"fallback_work":240156992,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":852.316162109375,"n":8,"spread_pct":1.4514963391058429},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.065692588686943,"wave_count":2},"fallback_work":241634752,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":851.7830200195312,"n":8,"spread_pct":1.90200172504067},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.2529947608709335,"wave_count":2},"fallback_work":249652160,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":848.3067321777344,"n":8,"spread_pct":1.5499338695122173},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.173737347126007,"wave_count":2},"fallback_work":252866368,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":849.1323852539062,"n":8,"spread_pct":0.9983190994081805},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.120545789599419,"wave_count":2},"fallback_work":254491904,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":850.0095520019531,"n":8,"spread_pct":1.8766376311330084},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.574233993887901,"wave_count":2},"fallback_work":263100416,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":920.7744750976562,"n":8,"spread_pct":1.0808453868624548},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.874250084161758,"wave_count":2},"fallback_work":272373440,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":920.7512512207031,"n":8,"spread_pct":0.9550174326974383},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.919037386775017,"wave_count":2},"fallback_work":275181184,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":921.4050598144531,"n":8,"spread_pct":1.7321727266888995},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.3193123787641525,"wave_count":2},"fallback_work":285932448,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":969.1688232421875,"n":8,"spread_pct":0.6676863870975136},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.653155967593193,"wave_count":1},"fallback_work":226594624,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":789.2707214355469,"n":8,"spread_pct":7.22970574403266},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":8.773799985647202,"wave_count":1},"fallback_work":288592416,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":960.188232421875,"n":8,"spread_pct":1.772119826155099},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.666019394993782,"wave_count":1},"fallback_work":227629696,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":790.6201477050781,"n":8,"spread_pct":18.18084126241974},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.65929152071476,"wave_count":1},"fallback_work":229921504,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":789.3110961914062,"n":8,"spread_pct":5.741577549090544},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.99162483215332,"wave_count":1},"fallback_work":298862848,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":957.4953918457031,"n":8,"spread_pct":1.3619478996936722},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":5.817342549562454,"wave_count":2},"fallback_work":496033696,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":1382.3704833984375,"n":8,"spread_pct":0.44024489113266163},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":6.469052284955978,"wave_count":1},"fallback_work":652343680,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":1545.7050170898438,"n":8,"spread_pct":0.6377377207374059},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":7.343715056777,"wave_count":1},"fallback_work":817631136,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":1656.4100952148438,"n":8,"spread_pct":0.6412779307188873},{"cp_plan":{"max_rank_tokens":16896,"summary_ms":8.194571360945702,"wave_count":1},"fallback_work":988164640,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":2006.5953979492188,"n":8,"spread_pct":0.5844252804145412},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":8.519984781742096,"wave_count":1},"fallback_work":1160988672,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":2188.4075927734375,"n":8,"spread_pct":0.8007726704907869},{"cp_plan":{"max_rank_tokens":20992,"summary_ms":8.911846205592155,"wave_count":1},"fallback_work":1331891616,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":2474.8387451171875,"n":8,"spread_pct":0.3810617593209469},{"cp_plan":{"max_rank_tokens":23040,"summary_ms":7.567413151264191,"wave_count":1},"fallback_work":1509075040,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":2735.3900146484375,"n":8,"spread_pct":0.4083662693959809},{"cp_plan":{"max_rank_tokens":26112,"summary_ms":7.316412404179573,"wave_count":2},"fallback_work":1684189600,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":3078.796142578125,"n":8,"spread_pct":0.35281013046885634},{"cp_plan":{"max_rank_tokens":27648,"summary_ms":9.48675349354744,"wave_count":1},"fallback_work":1861668576,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":3281.9976806640625,"n":8,"spread_pct":0.3549191265011228},{"cp_plan":{"max_rank_tokens":30720,"summary_ms":11.314840987324715,"wave_count":1},"fallback_work":2046758016,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":3634.784423828125,"n":8,"spread_pct":15.976072665395268},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.385989487171173,"wave_count":1},"fallback_work":313455808,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":959.0054931640625,"n":8,"spread_pct":1.697499533169649},{"cp_plan":{"max_rank_tokens":32768,"summary_ms":9.41513106226921,"wave_count":1},"fallback_work":2236169984,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":3910.7474365234375,"n":8,"spread_pct":0.20383406692101338},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.255276665091515,"wave_count":1},"fallback_work":234096896,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":846.4147338867188,"n":8,"spread_pct":1.3435793308733652},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.392713472247124,"wave_count":1},"fallback_work":236794368,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":845.2924194335938,"n":8,"spread_pct":1.1816867472663364},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.128858610987663,"wave_count":1},"fallback_work":237459360,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":844.6219787597656,"n":8,"spread_pct":1.91658121834665},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.187878385186195,"wave_count":1},"fallback_work":237865824,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":843.052978515625,"n":8,"spread_pct":1.210338691251019},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.408870056271553,"wave_count":1},"fallback_work":237016032,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":846.16259765625,"n":8,"spread_pct":76.45186708606929},{"cp_plan":{"max_rank_tokens":35328,"summary_ms":11.776013299822807,"wave_count":1},"fallback_work":2434596768,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":4232.327392578125,"n":8,"spread_pct":0.31553542410066493},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.398111745715141,"wave_count":1},"fallback_work":235794400,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":844.1544189453125,"n":8,"spread_pct":1.25303800420892},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.368437618017197,"wave_count":1},"fallback_work":230104384,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":791.5409545898438,"n":8,"spread_pct":1.893935100354263},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.852954462170601,"wave_count":1},"fallback_work":225744352,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":789.5918884277344,"n":8,"spread_pct":3.698701988575347},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.102053448557854,"wave_count":1},"fallback_work":321509600,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":961.8237915039062,"n":8,"spread_pct":1.0328507351673422},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.618436425924301,"wave_count":1},"fallback_work":301854752,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":966.2123107910156,"n":8,"spread_pct":1.0101688694663709},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.470441818237305,"wave_count":1},"fallback_work":289182240,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":914.9324340820312,"n":8,"spread_pct":1.2718998138323616},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.179600954055786,"wave_count":2},"fallback_work":280315040,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":925.8718566894531,"n":8,"spread_pct":1.5513249642728466},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.475141271948814,"wave_count":2},"fallback_work":272593184,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":929.2370910644531,"n":8,"spread_pct":2.124302540086303},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.786104291677475,"wave_count":2},"fallback_work":261804896,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":925.1901245117188,"n":8,"spread_pct":3.5402615660117895},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.560264155268669,"wave_count":2},"fallback_work":255449888,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":852.2909545898438,"n":8,"spread_pct":2.0335528631435196},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.345223471522331,"wave_count":2},"fallback_work":245696032,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":854.3939208984375,"n":8,"spread_pct":1.3540271416927232}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.12061470746994,"wave_count":2},"fallback_work":455684416,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1351.3250732421875,"n":8,"spread_pct":3.1942659724816185},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":9.386604651808739,"wave_count":1},"fallback_work":872117184,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":2018.7418212890625,"n":8,"spread_pct":0.5292325040258423},{"cp_plan":{"max_rank_tokens":20480,"summary_ms":6.69005885720253,"wave_count":1},"fallback_work":1291025120,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":2687.163818359375,"n":8,"spread_pct":0.896451163994624},{"cp_plan":{"max_rank_tokens":26624,"summary_ms":9.64735820889473,"wave_count":1},"fallback_work":1706977696,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":3512.5247802734375,"n":8,"spread_pct":0.4324994509915567},{"cp_plan":{"max_rank_tokens":31232,"summary_ms":7.02277198433876,"wave_count":1},"fallback_work":2121637152,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":4104.03125,"n":8,"spread_pct":5.601711636424552},{"cp_plan":{"max_rank_tokens":36864,"summary_ms":6.873214617371559,"wave_count":1},"fallback_work":2540212672,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":4854.937744140625,"n":8,"spread_pct":1.2328379637254174},{"cp_plan":{"max_rank_tokens":43008,"summary_ms":7.685493677854538,"wave_count":1},"fallback_work":2963812576,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":5626.320556640625,"n":8,"spread_pct":0.25875869679726665},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.545601412653923,"wave_count":2},"fallback_work":442049440,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":1265.2918090820312,"n":8,"spread_pct":1.4662907253602304},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.566950887441635,"wave_count":2},"fallback_work":450881056,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1358.420166015625,"n":8,"spread_pct":45.23843885315278},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.382295861840248,"wave_count":2},"fallback_work":442234880,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1262.9136352539062,"n":8,"spread_pct":4.9225924152615335},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.387978792190552,"wave_count":2},"fallback_work":441827136,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":1263.7350463867188,"n":8,"spread_pct":1.5253198074767804},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.609981879591942,"wave_count":2},"fallback_work":443269952,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1347.4647216796875,"n":8,"spread_pct":1.760340758573388},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.545860156416893,"wave_count":2},"fallback_work":444821600,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1347.3347778320312,"n":8,"spread_pct":3.0390763956239017},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.550572648644447,"wave_count":2},"fallback_work":446447776,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1354.4908447265625,"n":8,"spread_pct":1.7253622007015328},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.3966382294893265,"wave_count":2},"fallback_work":441717024,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":1265.0541381835938,"n":8,"spread_pct":2.120121021021876},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.487341433763504,"wave_count":2},"fallback_work":446152224,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1350.9620971679688,"n":8,"spread_pct":1.5528393822707458},{"cp_plan":{"max_rank_tokens":49280,"summary_ms":7.658068090677261,"wave_count":1},"fallback_work":3387487008,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":6467.94921875,"n":8,"spread_pct":8.182881938887363},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.926715210080147,"wave_count":2},"fallback_work":453541024,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1349.9404296875,"n":8,"spread_pct":2.083280584574406},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.9151072055101395,"wave_count":2},"fallback_work":453466496,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1351.6629028320312,"n":8,"spread_pct":1.9591120511504634},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.674993619322777,"wave_count":2},"fallback_work":450362560,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1349.90283203125,"n":8,"spread_pct":1.6787051730068823},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.66633976995945,"wave_count":2},"fallback_work":445115872,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":1353.0728759765625,"n":8,"spread_pct":2.513827349548775},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.474951282143593,"wave_count":2},"fallback_work":441938048,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":1269.5988159179688,"n":8,"spread_pct":35.481707776417856},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.43063522875309,"wave_count":2},"fallback_work":442048160,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":1269.0711669921875,"n":8,"spread_pct":2.0092858652332755}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.6056935787200928,"wave_count":2},"fallback_work":451915168,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":823.82763671875,"n":8,"spread_pct":1.2853996475815834},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":6.953936070203781,"wave_count":3},"fallback_work":884691520,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":1144.4226684570312,"n":8,"spread_pct":0.3327854109026235},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":10.39186678826809,"wave_count":1},"fallback_work":1322267360,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":1513.3679809570312,"n":8,"spread_pct":0.7261054548007845},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.6610662937164307,"wave_count":2},"fallback_work":450495264,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":820.4380493164062,"n":8,"spread_pct":1.2953895374792712},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":6.017034873366356,"wave_count":1},"fallback_work":1759786304,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1772.1957397460938,"n":8,"spread_pct":0.6441390765939325},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.545086994767189,"wave_count":2},"fallback_work":449018464,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":821.5866394042969,"n":8,"spread_pct":2.394086476767429},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.59916889667511,"wave_count":2},"fallback_work":449259904,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":822.7364196777344,"n":8,"spread_pct":1.0838880218040243}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.575960174202919,"wave_count":2},"fallback_work":440971136,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":802.9843444824219,"n":8,"spread_pct":1.4560703059657782},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":8.811984211206436,"wave_count":3},"fallback_work":876496512,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1163.1198120117188,"n":8,"spread_pct":0.6872805709111642}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.4652557224035263,"wave_count":2},"fallback_work":448364448,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":788.4344482421875,"n":8,"spread_pct":1.420228507994064}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":8192,"summary_ms":6.5457504242658615,"wave_count":1},"fallback_work":765853952,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":1080.5119018554688,"n":8,"spread_pct":1.5831797209382512},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":3.8496628403663635,"wave_count":2},"fallback_work":598404640,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":954.7117614746094,"n":8,"spread_pct":1.55706456288591},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.8666538894176483,"wave_count":2},"fallback_work":547770208,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":901.0537109375,"n":8,"spread_pct":3.2247740004920455},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.399925470352173,"wave_count":2},"fallback_work":812872864,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":1029.3515625,"n":8,"spread_pct":1.1721388616923578},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":4.747835919260979,"wave_count":1},"fallback_work":1017188256,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":1418.6249389648438,"n":8,"spread_pct":1.0510631895510132},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":5.168557167053223,"wave_count":2},"fallback_work":1302744480,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":1264.693115234375,"n":8,"spread_pct":0.43476263290312334},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":8.08854028582573,"wave_count":1},"fallback_work":1705115808,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":1810.1275024414062,"n":8,"spread_pct":1.0326218697703342},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":5.971914157271385,"wave_count":3},"fallback_work":578376832,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1140.6832885742188,"n":8,"spread_pct":1.1222778373643685},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":10.414140298962593,"wave_count":2},"fallback_work":537969408,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1027.1303100585938,"n":8,"spread_pct":0.8087944602083739}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7680,"summary_ms":6.7538972944021225,"wave_count":2},"fallback_work":621370112,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":1028.5375366210938,"n":8,"spread_pct":1.0123217679741086},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.016656756401062,"wave_count":4},"fallback_work":701523488,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":962.8426818847656,"n":8,"spread_pct":0.6609735797887248},{"cp_plan":{"max_rank_tokens":12288,"summary_ms":9.207237511873245,"wave_count":1},"fallback_work":1116853568,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":1515.9981079101562,"n":8,"spread_pct":0.7102310872005454},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.6430489271879196,"wave_count":2},"fallback_work":453994752,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":814.9979858398438,"n":8,"spread_pct":2.593274118850496}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.146155297756195,"wave_count":4},"fallback_work":484412736,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1075.6419677734375,"n":8,"spread_pct":0.6722797602585393},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":4.390805959701538,"wave_count":2},"fallback_work":925015904,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":1230.6832275390625,"n":8,"spread_pct":0.8641450096446482},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":8.564848452806473,"wave_count":2},"fallback_work":1357127584,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":1534.7628173828125,"n":8,"spread_pct":0.7552349015655013},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.371302530169487,"wave_count":1},"fallback_work":1806591776,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":1854.892578125,"n":8,"spread_pct":0.5025968590246445},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.771468997001648,"wave_count":2},"fallback_work":478229984,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":900.5856323242188,"n":8,"spread_pct":2.4572904661714063},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.242880269885063,"wave_count":2},"fallback_work":471695200,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":824.5026245117188,"n":8,"spread_pct":71.25853647414534},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":14.52893577516079,"wave_count":3},"fallback_work":2255703712,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2439.3902587890625,"n":8,"spread_pct":22.87111582389158},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.7198904901742935,"wave_count":2},"fallback_work":482512704,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":896.0457763671875,"n":8,"spread_pct":1.5130223705579269},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.2026471346616745,"wave_count":2},"fallback_work":471345920,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":818.3277893066406,"n":8,"spread_pct":0.6597137083714262},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.337502643465996,"wave_count":2},"fallback_work":471734816,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":822.5972290039062,"n":8,"spread_pct":1.2150014027151905}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":4096,"summary_ms":3.718087449669838,"wave_count":1},"fallback_work":337829344,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":630.4609375,"n":8,"spread_pct":3.9505697871101253},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6573726683855057,"wave_count":2},"fallback_work":241510624,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":620.7169494628906,"n":8,"spread_pct":10.32123802702042},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.640618175268173,"wave_count":2},"fallback_work":242987424,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":616.1768798828125,"n":8,"spread_pct":2.1509707053935823},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.699330613017082,"wave_count":2},"fallback_work":250868384,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":621.5257568359375,"n":8,"spread_pct":2.6588817251522574},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6711785942316055,"wave_count":2},"fallback_work":254080480,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":622.3079528808594,"n":8,"spread_pct":5.127132801063487},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6638136953115463,"wave_count":2},"fallback_work":255704960,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":621.7767944335938,"n":8,"spread_pct":4.469632502227553},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6630239337682724,"wave_count":1},"fallback_work":264176640,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":591.883056640625,"n":8,"spread_pct":108.92777371942121},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.695257008075714,"wave_count":2},"fallback_work":273443616,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":635.6343078613281,"n":8,"spread_pct":3.481408735593333},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.652408719062805,"wave_count":2},"fallback_work":276249536,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":637.74658203125,"n":8,"spread_pct":3.5041975951397477},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6078710108995438,"wave_count":1},"fallback_work":286862688,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":593.1228332519531,"n":8,"spread_pct":3.3803489695555786},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.9739739149808884,"wave_count":2},"fallback_work":228744672,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":619.1568298339844,"n":8,"spread_pct":2.6521782183244316},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.657367080450058,"wave_count":1},"fallback_work":289520928,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":592.2109375,"n":8,"spread_pct":1.755929019959632},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.9253029972314835,"wave_count":2},"fallback_work":229647808,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":615.9435424804688,"n":8,"spread_pct":3.198456065374644},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.77727672457695,"wave_count":2},"fallback_work":231675712,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":615.5985717773438,"n":8,"spread_pct":3.458379350239794},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6246348172426224,"wave_count":2},"fallback_work":299784576,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":647.6077880859375,"n":8,"spread_pct":6.118220169897363},{"cp_plan":{"max_rank_tokens":4608,"summary_ms":3.923315554857254,"wave_count":3},"fallback_work":496696160,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":763.7991333007812,"n":8,"spread_pct":2.725440973111498},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.152929410338402,"wave_count":1},"fallback_work":652904512,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":816.5429992675781,"n":8,"spread_pct":80.60326767185865},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.468284547328949,"wave_count":1},"fallback_work":818084704,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":936.6189270019531,"n":8,"spread_pct":0.5886717298669593},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":4.712244495749474,"wave_count":3},"fallback_work":988507424,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":1016.652587890625,"n":8,"spread_pct":1.1251416535529317},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":5.105271935462952,"wave_count":1},"fallback_work":1161219072,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":1125.183349609375,"n":8,"spread_pct":1.531499532085808},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.517268553376198,"wave_count":1},"fallback_work":1332011104,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":1229.986083984375,"n":8,"spread_pct":1.1390777691963692},{"cp_plan":{"max_rank_tokens":12288,"summary_ms":5.943248048424721,"wave_count":3},"fallback_work":1509079424,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":1461.7159423828125,"n":8,"spread_pct":0.52061163599576},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":10.00291295349598,"wave_count":3},"fallback_work":1684080224,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":1579.8189086914062,"n":8,"spread_pct":1.059112254702786},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.703289225697517,"wave_count":3},"fallback_work":1861443904,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":1655.862548828125,"n":8,"spread_pct":0.8604459799189463},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":7.061365991830826,"wave_count":1},"fallback_work":2046412992,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":1798.570068359375,"n":8,"spread_pct":0.5310954625276726},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.683304414153099,"wave_count":1},"fallback_work":314368032,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":591.84228515625,"n":8,"spread_pct":2.5062227073335133},{"cp_plan":{"max_rank_tokens":16896,"summary_ms":7.572893053293228,"wave_count":1},"fallback_work":2235701888,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":2010.522705078125,"n":8,"spread_pct":0.2154982677200923},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.7874169647693634,"wave_count":1},"fallback_work":235716992,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":551.0588989257812,"n":8,"spread_pct":5.324166707186282},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.7437789142131805,"wave_count":2},"fallback_work":238281472,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":619.34375,"n":8,"spread_pct":4.543728161738735},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.7244316190481186,"wave_count":2},"fallback_work":238946144,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":623.1560363769531,"n":8,"spread_pct":3.042795646550367},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.7335287779569626,"wave_count":2},"fallback_work":239352320,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":618.0516052246094,"n":8,"spread_pct":5.111060143000936},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.706475719809532,"wave_count":2},"fallback_work":238503104,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":617.7220764160156,"n":8,"spread_pct":6.345571826459449},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":6.370373070240021,"wave_count":1},"fallback_work":2433868640,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":2200.286865234375,"n":8,"spread_pct":0.5447953508370436},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.9474312216043472,"wave_count":1},"fallback_work":237807296,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":552.2117614746094,"n":8,"spread_pct":1.7081079587452248},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.060897976160049,"wave_count":2},"fallback_work":232252128,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":617.2633972167969,"n":8,"spread_pct":2.9221189028030423},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.010843113064766,"wave_count":2},"fallback_work":228026304,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":616.9897766113281,"n":8,"spread_pct":105.16195333740473},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":3.921598196029663,"wave_count":1},"fallback_work":322416704,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":623.2949523925781,"n":8,"spread_pct":2.337548886417629},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.749668598175049,"wave_count":2},"fallback_work":302905888,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":646.5205383300781,"n":8,"spread_pct":1.2745605006517526},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.772567957639694,"wave_count":1},"fallback_work":290372896,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":591.8774108886719,"n":8,"spread_pct":1.1715711749004454},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.811052069067955,"wave_count":2},"fallback_work":281642720,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":640.7678527832031,"n":8,"spread_pct":7.297648614083341},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.831705078482628,"wave_count":2},"fallback_work":274057120,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":632.4624633789062,"n":8,"spread_pct":5.968289751323673},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.8625728338956833,"wave_count":2},"fallback_work":263407104,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":641.6648559570312,"n":8,"spread_pct":2.785417762726413},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.469137638807297,"wave_count":2},"fallback_work":257187488,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":625.3593139648438,"n":8,"spread_pct":4.344679967678737},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.8925912231206894,"wave_count":2},"fallback_work":247571232,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":619.5173034667969,"n":8,"spread_pct":2.0336897755865504}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.811391070485115,"wave_count":2},"fallback_work":456044768,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":831.0694274902344,"n":8,"spread_pct":2.8806480848624783},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":6.154799833893776,"wave_count":3},"fallback_work":872207008,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":1069.8233642578125,"n":8,"spread_pct":0.82817272902888},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.465424180030823,"wave_count":2},"fallback_work":1290842944,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":1414.6309814453125,"n":8,"spread_pct":0.5556122680997462},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":6.230490282177925,"wave_count":1},"fallback_work":1706525280,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":1770.6443481445312,"n":8,"spread_pct":0.5530047893093738},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":11.72257587313652,"wave_count":3},"fallback_work":2120915360,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":2261.9266357421875,"n":8,"spread_pct":0.362920010104857},{"cp_plan":{"max_rank_tokens":19456,"summary_ms":8.103601634502411,"wave_count":1},"fallback_work":2539218848,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":2559.4158935546875,"n":8,"spread_pct":0.5972317578297242},{"cp_plan":{"max_rank_tokens":21504,"summary_ms":9.17300395667553,"wave_count":1},"fallback_work":2962543680,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":2876.644287109375,"n":8,"spread_pct":0.3809811559031074},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.866497427225113,"wave_count":2},"fallback_work":442943840,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":831.8007507324219,"n":8,"spread_pct":2.2188805459493675},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7544313818216324,"wave_count":2},"fallback_work":451375904,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":830.6304626464844,"n":8,"spread_pct":1.333788961528871},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.8114897906780243,"wave_count":2},"fallback_work":442997760,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":831.3747253417969,"n":8,"spread_pct":1.6005073251939148},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.441807046532631,"wave_count":2},"fallback_work":442852832,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":832.8067626953125,"n":8,"spread_pct":1.7046098318278167},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.803575411438942,"wave_count":2},"fallback_work":443900896,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":833.0275268554688,"n":8,"spread_pct":0.3030702013719927},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.688024520874023,"wave_count":2},"fallback_work":445451648,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":833.2631530761719,"n":8,"spread_pct":0.6650291508314403},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.681725054979324,"wave_count":2},"fallback_work":446945504,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":831.7687072753906,"n":8,"spread_pct":2.591582136925028},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7987325340509415,"wave_count":2},"fallback_work":442611616,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":831.4520263671875,"n":8,"spread_pct":2.268708189606761},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.669666290283203,"wave_count":2},"fallback_work":446650144,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":833.5131225585938,"n":8,"spread_pct":0.9365801621364063},{"cp_plan":{"max_rank_tokens":24576,"summary_ms":8.580336347222328,"wave_count":1},"fallback_work":3385811616,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3255.032958984375,"n":8,"spread_pct":18.830822010535815},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7905871868133545,"wave_count":2},"fallback_work":454034144,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":835.1889343261719,"n":8,"spread_pct":2.198595551631351},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.8035716861486435,"wave_count":2},"fallback_work":454090816,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":833.0250244140625,"n":8,"spread_pct":72.19658122538792},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.771575331687927,"wave_count":2},"fallback_work":451120160,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":831.1297912597656,"n":8,"spread_pct":1.653559437067618},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.831554368138313,"wave_count":2},"fallback_work":446008256,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":833.3493957519531,"n":8,"spread_pct":1.223362650273135},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.9238687604665756,"wave_count":2},"fallback_work":442963648,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":830.9179382324219,"n":8,"spread_pct":1.5695297099967132},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.89222614467144,"wave_count":2},"fallback_work":443205088,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":837.0217895507812,"n":8,"spread_pct":74.70826574613172}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12204,"summary_ms":0.0003203749656677246,"wave_count":0},"fallback_work":452204032,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":969.2489624023438,"n":8,"spread_pct":0.6166178899162006},{"cp_plan":{"max_rank_tokens":23926,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":885824256,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":1838.0963134765625,"n":8,"spread_pct":0.7636366413562243},{"cp_plan":{"max_rank_tokens":35778,"summary_ms":0.00018440186977386475,"wave_count":0},"fallback_work":1324253440,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":2684.0794677734375,"n":8,"spread_pct":0.25077338849559566},{"cp_plan":{"max_rank_tokens":12154,"summary_ms":0.00018440186977386475,"wave_count":0},"fallback_work":450583808,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":963.7941589355469,"n":8,"spread_pct":55.31438599688736},{"cp_plan":{"max_rank_tokens":47640,"summary_ms":0.0003501772880554199,"wave_count":0},"fallback_work":1762823168,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3591.6912841796875,"n":8,"spread_pct":0.4098543544051305},{"cp_plan":{"max_rank_tokens":12114,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":449104128,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":962.0737915039062,"n":8,"spread_pct":0.769364630251436},{"cp_plan":{"max_rank_tokens":12109,"summary_ms":0.000303611159324646,"wave_count":0},"fallback_work":449148544,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":961.0382385253906,"n":8,"spread_pct":0.6646597420665472}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11920,"summary_ms":0.00031478703022003174,"wave_count":0},"fallback_work":441436160,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":920.9605407714844,"n":8,"spread_pct":0.7082237130416211},{"cp_plan":{"max_rank_tokens":23728,"summary_ms":0.00033155083656311035,"wave_count":0},"fallback_work":878008320,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1778.7994384765625,"n":8,"spread_pct":0.6913568385782936}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12138,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":449139968,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":948.4682006835938,"n":8,"spread_pct":1.2028415214885382}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":20701,"summary_ms":0.00030547380447387695,"wave_count":0},"fallback_work":766656128,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":1621.314453125,"n":8,"spread_pct":34.87609518222218},{"cp_plan":{"max_rank_tokens":16154,"summary_ms":0.0003203749656677246,"wave_count":0},"fallback_work":598682880,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":1258.10546875,"n":8,"spread_pct":0.732816502367461},{"cp_plan":{"max_rank_tokens":14771,"summary_ms":0.00024400651454925537,"wave_count":0},"fallback_work":547752320,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1163.5088500976562,"n":8,"spread_pct":0.7976213938582052},{"cp_plan":{"max_rank_tokens":21963,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":813569408,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":1651.958740234375,"n":8,"spread_pct":0.5977826098072793},{"cp_plan":{"max_rank_tokens":27497,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":1018283136,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":1989.5869750976562,"n":8,"spread_pct":31.635689393934047},{"cp_plan":{"max_rank_tokens":35243,"summary_ms":0.0003073364496231079,"wave_count":0},"fallback_work":1304593792,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":2530.6654052734375,"n":8,"spread_pct":0.3970438581711411},{"cp_plan":{"max_rank_tokens":46153,"summary_ms":0.0004209578037261963,"wave_count":0},"fallback_work":1707947136,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":3378.9381103515625,"n":8,"spread_pct":0.34181758618533253},{"cp_plan":{"max_rank_tokens":15600,"summary_ms":0.00028312206268310547,"wave_count":0},"fallback_work":578418688,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1212.916259765625,"n":8,"spread_pct":1.2243098726778117},{"cp_plan":{"max_rank_tokens":14494,"summary_ms":0.0001862645149230957,"wave_count":0},"fallback_work":537734912,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1142.8333740234375,"n":8,"spread_pct":1.2329937831196043}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16800,"summary_ms":0.0002477318048477173,"wave_count":0},"fallback_work":622088192,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":1313.3069458007812,"n":8,"spread_pct":0.9321185110289103},{"cp_plan":{"max_rank_tokens":18971,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":702397824,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":1425.407958984375,"n":8,"spread_pct":1.0592328025608075},{"cp_plan":{"max_rank_tokens":30232,"summary_ms":0.0004936009645462036,"wave_count":0},"fallback_work":1118735360,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":2243.1314697265625,"n":8,"spread_pct":0.34867724732395705},{"cp_plan":{"max_rank_tokens":12255,"summary_ms":0.0002384185791015625,"wave_count":0},"fallback_work":454188928,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":954.0795288085938,"n":8,"spread_pct":0.6640890312269877}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13078,"summary_ms":0.0004209578037261963,"wave_count":0},"fallback_work":484666112,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1052.5687255859375,"n":8,"spread_pct":0.4645559371993018},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":926128640,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":1943.0556030273438,"n":8,"spread_pct":24.106594802837115},{"cp_plan":{"max_rank_tokens":36716,"summary_ms":0.00025890767574310303,"wave_count":0},"fallback_work":1359083008,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":2773.88623046875,"n":8,"spread_pct":0.4257145718852883},{"cp_plan":{"max_rank_tokens":48890,"summary_ms":0.00027008354663848877,"wave_count":0},"fallback_work":1809423616,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":3686.7628173828125,"n":8,"spread_pct":0.46012906248041313},{"cp_plan":{"max_rank_tokens":12899,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":478273920,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1050.3883666992188,"n":8,"spread_pct":0.7337432100977811},{"cp_plan":{"max_rank_tokens":12722,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":471726336,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1025.9637451171875,"n":8,"spread_pct":0.8213801928564928},{"cp_plan":{"max_rank_tokens":61066,"summary_ms":0.0003110617399215698,"wave_count":0},"fallback_work":2259608832,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":4657.23046875,"n":8,"spread_pct":0.21256010689024377},{"cp_plan":{"max_rank_tokens":13015,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":482564992,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1046.1217651367188,"n":8,"spread_pct":0.9115233856169485},{"cp_plan":{"max_rank_tokens":12701,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":471178880,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1027.9227294921875,"n":8,"spread_pct":0.5896985574788353},{"cp_plan":{"max_rank_tokens":12700,"summary_ms":0.00025331974029541016,"wave_count":0},"fallback_work":471371264,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1026.5326538085938,"n":8,"spread_pct":0.4869398136397017}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9058,"summary_ms":0.00033527612686157227,"wave_count":0},"fallback_work":337006848,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":726.0315551757812,"n":8,"spread_pct":1.6182440444340165},{"cp_plan":{"max_rank_tokens":6403,"summary_ms":0.00021047890186309814,"wave_count":0},"fallback_work":239710592,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":541.1708984375,"n":8,"spread_pct":0.9936002122134622},{"cp_plan":{"max_rank_tokens":6443,"summary_ms":0.00019744038581848145,"wave_count":0},"fallback_work":241190272,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":543.9729919433594,"n":8,"spread_pct":0.8119537453968203},{"cp_plan":{"max_rank_tokens":6668,"summary_ms":0.00023469328880310059,"wave_count":0},"fallback_work":249284096,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":557.0201416015625,"n":8,"spread_pct":0.9119111553109953},{"cp_plan":{"max_rank_tokens":6755,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":252502400,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":569.2752990722656,"n":8,"spread_pct":1.1847531145175034},{"cp_plan":{"max_rank_tokens":6799,"summary_ms":0.0001862645149230957,"wave_count":0},"fallback_work":254130048,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":564.9448852539062,"n":8,"spread_pct":0.80165884890294},{"cp_plan":{"max_rank_tokens":7040,"summary_ms":0.00024028122425079346,"wave_count":0},"fallback_work":262815744,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":577.96337890625,"n":8,"spread_pct":0.9796868888177354},{"cp_plan":{"max_rank_tokens":7291,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":272100736,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":599.0441589355469,"n":8,"spread_pct":0.8384328488799933},{"cp_plan":{"max_rank_tokens":7367,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":274912128,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":600.8201293945312,"n":8,"spread_pct":1.0394727178030274},{"cp_plan":{"max_rank_tokens":7666,"summary_ms":0.00021792948246002197,"wave_count":0},"fallback_work":285743360,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":622.0117492675781,"n":8,"spread_pct":0.7476371033948889},{"cp_plan":{"max_rank_tokens":5988,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":225735168,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":516.6092224121094,"n":8,"spread_pct":2.0885382665935186},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":288406784,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":635.0707702636719,"n":8,"spread_pct":0.7105142100199043},{"cp_plan":{"max_rank_tokens":6024,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":226837504,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":512.4486541748047,"n":8,"spread_pct":4.373592235669333},{"cp_plan":{"max_rank_tokens":6102,"summary_ms":0.00018440186977386475,"wave_count":0},"fallback_work":229264128,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":523.59228515625,"n":8,"spread_pct":9.980758954674783},{"cp_plan":{"max_rank_tokens":8016,"summary_ms":0.0002514570951461792,"wave_count":0},"fallback_work":298690560,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":645.3721923828125,"n":8,"spread_pct":0.9921713885046575},{"cp_plan":{"max_rank_tokens":13361,"summary_ms":0.00024400651454925537,"wave_count":0},"fallback_work":496183424,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":994.4617919921875,"n":8,"spread_pct":0.6735483661896555},{"cp_plan":{"max_rank_tokens":17592,"summary_ms":0.00023655593395233154,"wave_count":0},"fallback_work":652696576,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":1253.8218383789062,"n":8,"spread_pct":0.733567685168837},{"cp_plan":{"max_rank_tokens":22066,"summary_ms":0.00017881393432617188,"wave_count":0},"fallback_work":818198784,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":1562.6456298828125,"n":8,"spread_pct":41.907390975200514},{"cp_plan":{"max_rank_tokens":26682,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":988953856,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":1858.748779296875,"n":8,"spread_pct":0.42414494902760713},{"cp_plan":{"max_rank_tokens":31360,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":1162002432,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":2142.91552734375,"n":8,"spread_pct":0.36067702191838064},{"cp_plan":{"max_rank_tokens":35986,"summary_ms":0.00024214386940002441,"wave_count":0},"fallback_work":1333127424,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":2451.5404052734375,"n":8,"spread_pct":0.35578316188417686},{"cp_plan":{"max_rank_tokens":40782,"summary_ms":0.00018067657947540283,"wave_count":0},"fallback_work":1510541056,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":2767.7039794921875,"n":8,"spread_pct":0.6211785270350488},{"cp_plan":{"max_rank_tokens":45522,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":1685883136,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":3074.334228515625,"n":8,"spread_pct":0.3205486325045128},{"cp_plan":{"max_rank_tokens":50326,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":1863592704,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":3407.5245361328125,"n":8,"spread_pct":0.6188266496183821},{"cp_plan":{"max_rank_tokens":55336,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":2048922624,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":3669.828369140625,"n":8,"spread_pct":0.4163424334188398},{"cp_plan":{"max_rank_tokens":8411,"summary_ms":0.00022351741790771484,"wave_count":0},"fallback_work":313302400,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":676.5453491210938,"n":8,"spread_pct":0.8529463479651214},{"cp_plan":{"max_rank_tokens":60463,"summary_ms":0.0001825392246246338,"wave_count":0},"fallback_work":2238580608,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":3997.503173828125,"n":8,"spread_pct":0.3023799096274267},{"cp_plan":{"max_rank_tokens":6223,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":233510784,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":524.8482971191406,"n":8,"spread_pct":2.0661429732096255},{"cp_plan":{"max_rank_tokens":6304,"summary_ms":0.00017881393432617188,"wave_count":0},"fallback_work":236277760,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":530.8934936523438,"n":8,"spread_pct":1.2699813834168816},{"cp_plan":{"max_rank_tokens":6322,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":236943616,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":533.7963256835938,"n":8,"spread_pct":0.858282778122645},{"cp_plan":{"max_rank_tokens":6333,"summary_ms":0.00019371509552001953,"wave_count":0},"fallback_work":237350528,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":534.5630187988281,"n":8,"spread_pct":0.6618308239842284},{"cp_plan":{"max_rank_tokens":6310,"summary_ms":0.00022351741790771484,"wave_count":0},"fallback_work":236499712,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":533.3706970214844,"n":8,"spread_pct":1.0475302204418244},{"cp_plan":{"max_rank_tokens":65842,"summary_ms":0.0005327165126800537,"wave_count":0},"fallback_work":2437331200,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":4478.116943359375,"n":8,"spread_pct":0.23086442442243144},{"cp_plan":{"max_rank_tokens":6245,"summary_ms":0.00022351741790771484,"wave_count":0},"fallback_work":235012736,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":533.6030883789062,"n":8,"spread_pct":0.8546588164807236},{"cp_plan":{"max_rank_tokens":6083,"summary_ms":0.00021420419216156006,"wave_count":0},"fallback_work":229249408,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":521.0016784667969,"n":8,"spread_pct":1.986002610211008},{"cp_plan":{"max_rank_tokens":5957,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":224817792,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":515.2049560546875,"n":8,"spread_pct":1.7258969724152653},{"cp_plan":{"max_rank_tokens":8629,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":321366656,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":696.0893249511719,"n":8,"spread_pct":1.0479250819451846},{"cp_plan":{"max_rank_tokens":8089,"summary_ms":0.00024586915969848633,"wave_count":0},"fallback_work":301620352,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":652.7850952148438,"n":8,"spread_pct":0.7592449382346206},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":288865536,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":634.7803039550781,"n":8,"spread_pct":1.891446644911721},{"cp_plan":{"max_rank_tokens":7490,"summary_ms":0.00018067657947540283,"wave_count":0},"fallback_work":279920896,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":616.6976318359375,"n":8,"spread_pct":0.6797523429890061},{"cp_plan":{"max_rank_tokens":7273,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":272123008,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":598.0330810546875,"n":8,"spread_pct":0.5353854708030292},{"cp_plan":{"max_rank_tokens":6973,"summary_ms":0.00021420419216156006,"wave_count":0},"fallback_work":261254784,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":576.1979064941406,"n":8,"spread_pct":0.6508180536123023},{"cp_plan":{"max_rank_tokens":6793,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":254825600,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":567.3092041015625,"n":8,"spread_pct":0.8459251490673166},{"cp_plan":{"max_rank_tokens":6521,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":244993152,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":547.0054931640625,"n":8,"spread_pct":1.167288430506203}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12291,"summary_ms":0.0002384185791015625,"wave_count":0},"fallback_work":455946624,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1019.2242431640625,"n":8,"spread_pct":0.25238083546470236},{"cp_plan":{"max_rank_tokens":23563,"summary_ms":0.00023655593395233154,"wave_count":0},"fallback_work":872920448,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":1837.8086547851562,"n":8,"spread_pct":1.0587480704921184},{"cp_plan":{"max_rank_tokens":34902,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":1292372736,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":2643.193115234375,"n":8,"spread_pct":0.2664845459560917},{"cp_plan":{"max_rank_tokens":46161,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":1708865664,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":3481.859130859375,"n":8,"spread_pct":0.3461860599318664},{"cp_plan":{"max_rank_tokens":57385,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":2124063872,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":4288.242431640625,"n":8,"spread_pct":0.5064720650994233},{"cp_plan":{"max_rank_tokens":68715,"summary_ms":0.00018067657947540283,"wave_count":0},"fallback_work":2543183232,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":5146.8212890625,"n":8,"spread_pct":0.722922313434889},{"cp_plan":{"max_rank_tokens":80181,"summary_ms":0.00035762786865234375,"wave_count":0},"fallback_work":2967333504,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":5957.9443359375,"n":8,"spread_pct":0.35833032149990124},{"cp_plan":{"max_rank_tokens":11890,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":442030336,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":988.5406188964844,"n":8,"spread_pct":0.6793301128140808},{"cp_plan":{"max_rank_tokens":12153,"summary_ms":0.00022351741790771484,"wave_count":0},"fallback_work":451071104,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1002.8460693359375,"n":8,"spread_pct":0.8856385908513551},{"cp_plan":{"max_rank_tokens":11903,"summary_ms":0.00024959444999694824,"wave_count":0},"fallback_work":442281856,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":988.2816162109375,"n":8,"spread_pct":0.7759507994253461},{"cp_plan":{"max_rank_tokens":11876,"summary_ms":0.0002216547727584839,"wave_count":0},"fallback_work":441741824,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":987.2149658203125,"n":8,"spread_pct":0.747663146935336},{"cp_plan":{"max_rank_tokens":11939,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":443384192,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":990.0700988769531,"n":8,"spread_pct":0.9198887993530238},{"cp_plan":{"max_rank_tokens":11981,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":444937856,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":989.3328247070312,"n":8,"spread_pct":1.4061927313137472},{"cp_plan":{"max_rank_tokens":12033,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":446632064,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":993.4608459472656,"n":8,"spread_pct":0.5421806567451191},{"cp_plan":{"max_rank_tokens":11881,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":441697408,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":987.4778442382812,"n":8,"spread_pct":0.7891172894565908},{"cp_plan":{"max_rank_tokens":12025,"summary_ms":0.00023283064365386963,"wave_count":0},"fallback_work":446336128,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":994.3730163574219,"n":8,"spread_pct":0.781705774512742},{"cp_plan":{"max_rank_tokens":91657,"summary_ms":0.0003520399332046509,"wave_count":0},"fallback_work":3391624320,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":6922.4423828125,"n":8,"spread_pct":0.5997533215745153},{"cp_plan":{"max_rank_tokens":12225,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":453734528,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1013.447021484375,"n":8,"spread_pct":1.761361323164411},{"cp_plan":{"max_rank_tokens":12215,"summary_ms":0.000303611159324646,"wave_count":0},"fallback_work":453593984,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":999.7807006835938,"n":8,"spread_pct":0.986281964788787},{"cp_plan":{"max_rank_tokens":12123,"summary_ms":0.00022351741790771484,"wave_count":0},"fallback_work":450420096,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1004.2945556640625,"n":8,"spread_pct":0.514823971904836},{"cp_plan":{"max_rank_tokens":11973,"summary_ms":0.0002905726432800293,"wave_count":0},"fallback_work":445100672,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":990.9570007324219,"n":8,"spread_pct":1.0030402185239637},{"cp_plan":{"max_rank_tokens":11879,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":441852800,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":977.5636901855469,"n":8,"spread_pct":1.0334842026176476},{"cp_plan":{"max_rank_tokens":11874,"summary_ms":0.0001825392246246338,"wave_count":0},"fallback_work":441897216,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":988.4277954101562,"n":8,"spread_pct":0.3848428848958353}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":74719,"summary_ms":0.0005681067705154419,"wave_count":0},"fallback_work":2768428928,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":7659.652587890625,"n":8,"spread_pct":0.6375036706422914},{"cp_plan":{"max_rank_tokens":14302,"summary_ms":0.0002905726432800293,"wave_count":0},"fallback_work":533712640,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1541.3911743164062,"n":8,"spread_pct":0.4880068636631765},{"cp_plan":{"max_rank_tokens":106432,"summary_ms":0.0012051314115524292,"wave_count":0},"fallback_work":3941326848,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":10777.826171875,"n":8,"spread_pct":0.41528586886378954},{"cp_plan":{"max_rank_tokens":13279,"summary_ms":0.0003334134817123413,"wave_count":0},"fallback_work":496099200,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1461.7353515625,"n":8,"spread_pct":0.9248124520231932}],"cell":"cal-grpo-g16|Qwen/Qwen3-8B|L36|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":38879,"summary_ms":23.41325208544731,"wave_count":1},"fallback_work":2765956608,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":4564.162353515625,"n":8,"spread_pct":0.8488773541361134},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":12.044733390212059,"wave_count":1},"fallback_work":534205792,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1180.1815795898438,"n":8,"spread_pct":2.0073649824331126},{"cp_plan":{"max_rank_tokens":55296,"summary_ms":7.893064990639687,"wave_count":1},"fallback_work":3937266688,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":6135.08251953125,"n":8,"spread_pct":0.31658663922883956},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":12.488095089793205,"wave_count":1},"fallback_work":496707072,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1079.964111328125,"n":8,"spread_pct":0.9419846082305705}],"cell":"cal-grpo-g16|Qwen/Qwen3-8B|L36|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19456,"summary_ms":7.773827761411667,"wave_count":1},"fallback_work":2766391552,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":2305.4232177734375,"n":8,"spread_pct":0.9068821091910966},{"cp_plan":{"max_rank_tokens":4608,"summary_ms":4.497705027461052,"wave_count":3},"fallback_work":536221952,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":754.0105895996094,"n":8,"spread_pct":2.965093197310492},{"cp_plan":{"max_rank_tokens":26880,"summary_ms":8.584858849644661,"wave_count":1},"fallback_work":3936809472,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":3046.0059814453125,"n":8,"spread_pct":1.0105205887849182},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":4.625018686056137,"wave_count":1},"fallback_work":498878720,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":665.72705078125,"n":8,"spread_pct":95.52267514736957}],"cell":"cal-grpo-g16|Qwen/Qwen3-8B|L36|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":74719,"summary_ms":0.00041909515857696533,"wave_count":0},"fallback_work":2768428928,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":4745.361083984375,"n":8,"spread_pct":0.3687194715603988},{"cp_plan":{"max_rank_tokens":14302,"summary_ms":0.0003632158041000366,"wave_count":0},"fallback_work":533712640,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":997.5610046386719,"n":8,"spread_pct":0.6181703121801682},{"cp_plan":{"max_rank_tokens":106432,"summary_ms":0.0006034970283508301,"wave_count":0},"fallback_work":3941326848,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":6592.65869140625,"n":8,"spread_pct":0.6000845074699492},{"cp_plan":{"max_rank_tokens":13279,"summary_ms":0.0002942979335784912,"wave_count":0},"fallback_work":496099200,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":944.9146423339844,"n":8,"spread_pct":0.4460942578396002}],"cell":"cal-grpo-g16|Qwen/Qwen3-8B|L36|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":58351,"summary_ms":0.00045262277126312256,"wave_count":0},"fallback_work":2160846720,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":6171.705078125,"n":8,"spread_pct":0.4441577980153251},{"cp_plan":{"max_rank_tokens":24556,"summary_ms":0.0005327165126800537,"wave_count":0},"fallback_work":911390208,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":2564.322509765625,"n":8,"spread_pct":0.22123175542547152},{"cp_plan":{"max_rank_tokens":73696,"summary_ms":0.0010114163160324097,"wave_count":0},"fallback_work":2728259584,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":7402.047119140625,"n":8,"spread_pct":0.18547577021968395},{"cp_plan":{"max_rank_tokens":21487,"summary_ms":0.0003129243850708008,"wave_count":0},"fallback_work":798091136,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":2344.0882568359375,"n":8,"spread_pct":0.5733026759523404}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":31727,"summary_ms":15.363626182079315,"wave_count":1},"fallback_work":2158635776,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3836.679443359375,"n":8,"spread_pct":0.8225751141947532},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":13.310056179761887,"wave_count":1},"fallback_work":910997952,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1759.8986206054688,"n":8,"spread_pct":0.6828909538225452},{"cp_plan":{"max_rank_tokens":37888,"summary_ms":9.111924096941948,"wave_count":1},"fallback_work":2725246464,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":4257.857177734375,"n":8,"spread_pct":10.03474562164858},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":15.89450053870678,"wave_count":1},"fallback_work":797911808,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1675.2777709960938,"n":8,"spread_pct":0.9745643641258622}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":15360,"summary_ms":7.388802245259285,"wave_count":1},"fallback_work":2158414976,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1834.3550415039062,"n":8,"spread_pct":0.7423619420901786},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.65763546526432,"wave_count":1},"fallback_work":911981472,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":870.2941284179688,"n":8,"spread_pct":67.89796322983983},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":6.948145106434822,"wave_count":1},"fallback_work":2724526336,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":2191.561767578125,"n":8,"spread_pct":0.8199730032527771},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.635039716959,"wave_count":1},"fallback_work":799100032,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":789.8424377441406,"n":8,"spread_pct":74.61327770968339}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":58351,"summary_ms":0.0004116445779800415,"wave_count":0},"fallback_work":2160846720,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3780.7685546875,"n":8,"spread_pct":0.5836745226083797},{"cp_plan":{"max_rank_tokens":24556,"summary_ms":0.00019744038581848145,"wave_count":0},"fallback_work":911390208,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1626.5062866210938,"n":8,"spread_pct":0.5352160232783021},{"cp_plan":{"max_rank_tokens":73696,"summary_ms":0.0004637986421585083,"wave_count":0},"fallback_work":2728259584,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":4585.357177734375,"n":8,"spread_pct":0.27427874577186434},{"cp_plan":{"max_rank_tokens":21487,"summary_ms":0.0002905726432800293,"wave_count":0},"fallback_work":798091136,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1470.3208618164062,"n":8,"spread_pct":1.1894857963787648}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":37871,"summary_ms":0.0010691583156585693,"wave_count":0},"fallback_work":1403250560,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3877.773193359375,"n":8,"spread_pct":0.40080942576441164},{"cp_plan":{"max_rank_tokens":10222,"summary_ms":0.0003073364496231079,"wave_count":0},"fallback_work":380688128,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1074.3084716796875,"n":8,"spread_pct":1.2079104756591352},{"cp_plan":{"max_rank_tokens":53216,"summary_ms":0.001169741153717041,"wave_count":0},"fallback_work":1970663424,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":5205.493896484375,"n":8,"spread_pct":0.7467133371100798},{"cp_plan":{"max_rank_tokens":9199,"summary_ms":0.0006146728992462158,"wave_count":0},"fallback_work":343074688,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1008.2011108398438,"n":8,"spread_pct":0.7754277405266513}],"cell":"cal-grpo-g8|Qwen/Qwen3-8B|L36|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":21487,"summary_ms":9.327063336968422,"wave_count":1},"fallback_work":1402022656,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":2525.421875,"n":8,"spread_pct":0.9517186948962735},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.427798420190811,"wave_count":1},"fallback_work":380852832,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":850.4678649902344,"n":8,"spread_pct":8.164666658126446},{"cp_plan":{"max_rank_tokens":28672,"summary_ms":10.7573252171278,"wave_count":1},"fallback_work":1968633344,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":3174.2752685546875,"n":8,"spread_pct":0.6285499085263488},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":8.040836080908775,"wave_count":1},"fallback_work":343354112,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":760.2027893066406,"n":8,"spread_pct":4.398112027843876}],"cell":"cal-grpo-g8|Qwen/Qwen3-8B|L36|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.975957959890366,"wave_count":1},"fallback_work":1402293376,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1197.1997680664062,"n":8,"spread_pct":8.607335201604942},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.8061626255512238,"wave_count":3},"fallback_work":381918336,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":685.5083618164062,"n":8,"spread_pct":98.01383072713575},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":6.819657981395721,"wave_count":1},"fallback_work":1968404736,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":1544.866455078125,"n":8,"spread_pct":2.1800323747391066},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.946090117096901,"wave_count":1},"fallback_work":344575104,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":579.797607421875,"n":8,"spread_pct":35.0810850870223}],"cell":"cal-grpo-g8|Qwen/Qwen3-8B|L36|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":37871,"summary_ms":0.000607222318649292,"wave_count":0},"fallback_work":1403250560,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":2419.3912353515625,"n":8,"spread_pct":0.9272008949904421},{"cp_plan":{"max_rank_tokens":10222,"summary_ms":0.0003203749656677246,"wave_count":0},"fallback_work":380688128,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":707.8462524414062,"n":8,"spread_pct":1.5038517841097017},{"cp_plan":{"max_rank_tokens":53216,"summary_ms":0.0008288770914077759,"wave_count":0},"fallback_work":1970663424,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":3250.0491943359375,"n":8,"spread_pct":0.4746021910801155},{"cp_plan":{"max_rank_tokens":9199,"summary_ms":0.0004433095455169678,"wave_count":0},"fallback_work":343074688,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":648.9169616699219,"n":8,"spread_pct":0.8135832922108468}],"cell":"cal-grpo-g8|Qwen/Qwen3-8B|L36|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":28852,"summary_ms":0.00038929283618927,"wave_count":0},"fallback_work":1070307840,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":2990.05029296875,"n":8,"spread_pct":0.3757089553030943},{"cp_plan":{"max_rank_tokens":33460,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":1240308224,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":3399.66015625,"n":8,"spread_pct":0.26865334655463036},{"cp_plan":{"max_rank_tokens":29364,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":1089018368,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":3050.513916015625,"n":8,"spread_pct":0.34255562567302433},{"cp_plan":{"max_rank_tokens":36532,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":1353718272,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":3686.3582763671875,"n":8,"spread_pct":0.22850030334140464},{"cp_plan":{"max_rank_tokens":48820,"summary_ms":0.0008400529623031616,"wave_count":0},"fallback_work":1808046592,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":5070.053955078125,"n":8,"spread_pct":0.6459882560459915}],"cell":"cal-hetero2|Qwen/Qwen3-8B|L36|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":18432,"summary_ms":5.744447931647301,"wave_count":1},"fallback_work":1069709376,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":2120.731689453125,"n":8,"spread_pct":0.41121199670708064},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":4.2032599449157715,"wave_count":1},"fallback_work":1239357504,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":2519.5966796875,"n":8,"spread_pct":0.5874459849298487},{"cp_plan":{"max_rank_tokens":18944,"summary_ms":5.621816962957382,"wave_count":1},"fallback_work":1088329792,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":2133.9005126953125,"n":8,"spread_pct":0.6104601058308567},{"cp_plan":{"max_rank_tokens":25088,"summary_ms":5.2460357546806335,"wave_count":1},"fallback_work":1352554560,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":2752.303955078125,"n":8,"spread_pct":1.9206121433951333},{"cp_plan":{"max_rank_tokens":30720,"summary_ms":6.3704196363687515,"wave_count":1},"fallback_work":1806227520,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":3427.04296875,"n":8,"spread_pct":0.5991379457736774}],"cell":"cal-hetero2|Qwen/Qwen3-8B|L36|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13312,"summary_ms":12.263286858797073,"wave_count":3},"fallback_work":1070589792,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":1508.7251586914062,"n":8,"spread_pct":37.98126886684885},{"cp_plan":{"max_rank_tokens":16384,"summary_ms":5.343746393918991,"wave_count":2},"fallback_work":1239865184,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":1805.6469116210938,"n":8,"spread_pct":0.5771755741730781},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":12.404600158333778,"wave_count":3},"fallback_work":1089066848,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":1563.0029296875,"n":8,"spread_pct":0.9508501950934544},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":5.580238997936249,"wave_count":2},"fallback_work":1352857440,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":1942.12841796875,"n":8,"spread_pct":0.5725863040383117},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":6.559835746884346,"wave_count":1},"fallback_work":1806104416,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":2476.5184326171875,"n":8,"spread_pct":44.63757463349974}],"cell":"cal-hetero2|Qwen/Qwen3-8B|L36|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":28852,"summary_ms":0.00040978193283081055,"wave_count":0},"fallback_work":1070307840,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":1890.7767333984375,"n":8,"spread_pct":0.5184888635293539},{"cp_plan":{"max_rank_tokens":33460,"summary_ms":0.00018440186977386475,"wave_count":0},"fallback_work":1240308224,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":2154.748046875,"n":8,"spread_pct":0.6553138010081297},{"cp_plan":{"max_rank_tokens":29364,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":1089018368,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":1926.5624389648438,"n":8,"spread_pct":0.682614833061348},{"cp_plan":{"max_rank_tokens":36532,"summary_ms":0.00019371509552001953,"wave_count":0},"fallback_work":1353718272,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":2334.9818115234375,"n":8,"spread_pct":0.3949149137154828},{"cp_plan":{"max_rank_tokens":48820,"summary_ms":0.0004507601261138916,"wave_count":0},"fallback_work":1808046592,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":3189.479736328125,"n":8,"spread_pct":0.3978763496232311}],"cell":"cal-hetero2|Qwen/Qwen3-8B|L36|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16756,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":621738496,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1704.7952880859375,"n":8,"spread_pct":3.804110404052562},{"cp_plan":{"max_rank_tokens":17524,"summary_ms":0.00035390257835388184,"wave_count":0},"fallback_work":649918976,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1772.238037109375,"n":8,"spread_pct":0.8445345812772321},{"cp_plan":{"max_rank_tokens":25204,"summary_ms":0.000867992639541626,"wave_count":0},"fallback_work":933788160,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":2550.6358642578125,"n":8,"spread_pct":0.4812678593999113}],"cell":"cal-hetero3|Qwen/Qwen3-8B|L36|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9588,"summary_ms":5.2556488662958145,"wave_count":1},"fallback_work":621425728,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1077.6942749023438,"n":8,"spread_pct":1.6697907446379592},{"cp_plan":{"max_rank_tokens":10356,"summary_ms":5.516912788152695,"wave_count":1},"fallback_work":649503808,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1150.4434204101562,"n":8,"spread_pct":165.33975918443895},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":6.690051406621933,"wave_count":1},"fallback_work":932938816,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1562.8114624023438,"n":8,"spread_pct":38.081252787774375}],"cell":"cal-hetero3|Qwen/Qwen3-8B|L36|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":10.05454733967781,"wave_count":3},"fallback_work":622006624,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":751.4710998535156,"n":8,"spread_pct":2.6403451400766036},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":9.686484932899475,"wave_count":3},"fallback_work":649935200,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":794.0420837402344,"n":8,"spread_pct":1.6985322091929833},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":4.443114623427391,"wave_count":3},"fallback_work":933054816,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":868.0092163085938,"n":8,"spread_pct":1.7445673486616673}],"cell":"cal-hetero3|Qwen/Qwen3-8B|L36|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16756,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":621738496,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1105.1591796875,"n":8,"spread_pct":0.9978390591711637},{"cp_plan":{"max_rank_tokens":17524,"summary_ms":0.0005047768354415894,"wave_count":0},"fallback_work":649918976,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1143.096923828125,"n":8,"spread_pct":0.5622126705288731},{"cp_plan":{"max_rank_tokens":25204,"summary_ms":0.00045634806156158447,"wave_count":0},"fallback_work":933788160,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1635.9453735351562,"n":8,"spread_pct":1.010329724341331}],"cell":"cal-hetero3|Qwen/Qwen3-8B|L36|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19380,"summary_ms":0.00039301812648773193,"wave_count":0},"fallback_work":719165952,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1969.3589477539062,"n":8,"spread_pct":0.6747416946339483},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":927046144,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":2480.570068359375,"n":8,"spread_pct":0.42715774896989933},{"cp_plan":{"max_rank_tokens":20916,"summary_ms":0.0002905726432800293,"wave_count":0},"fallback_work":775756288,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":2083.35595703125,"n":8,"spread_pct":0.8105068234024647},{"cp_plan":{"max_rank_tokens":29108,"summary_ms":0.0009201467037200928,"wave_count":0},"fallback_work":1078336000,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":2953.4615478515625,"n":8,"spread_pct":0.4603063382698987}],"cell":"cal-hetero|Qwen/Qwen3-8B|L36|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11264,"summary_ms":10.549169033765793,"wave_count":1},"fallback_work":718825536,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1277.2403564453125,"n":8,"spread_pct":0.9687041755004443},{"cp_plan":{"max_rank_tokens":15872,"summary_ms":8.531078696250916,"wave_count":1},"fallback_work":926304320,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1760.3668823242188,"n":8,"spread_pct":0.40096594355195864},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":11.645082384347916,"wave_count":1},"fallback_work":775276608,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1437.7265625,"n":8,"spread_pct":0.8372067038347216},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":11.614259332418442,"wave_count":1},"fallback_work":1077332032,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1922.232421875,"n":8,"spread_pct":0.3352717431615868}],"cell":"cal-hetero|Qwen/Qwen3-8B|L36|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6580,"summary_ms":4.230327904224396,"wave_count":3},"fallback_work":719540064,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":866.2955017089844,"n":8,"spread_pct":2.6431396512128833},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":4.491845145821571,"wave_count":2},"fallback_work":926621536,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1095.460205078125,"n":8,"spread_pct":1.250366335673796},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.252634942531586,"wave_count":1},"fallback_work":775823200,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":865.4682922363281,"n":8,"spread_pct":2.9361841897638774},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":9.65229980647564,"wave_count":3},"fallback_work":1077419872,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1215.5215454101562,"n":8,"spread_pct":0.9325383923408891}],"cell":"cal-hetero|Qwen/Qwen3-8B|L36|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19380,"summary_ms":0.00042282044887542725,"wave_count":0},"fallback_work":719165952,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1258.6956787109375,"n":8,"spread_pct":0.5397510579479567},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.00038370490074157715,"wave_count":0},"fallback_work":927046144,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1590.656005859375,"n":8,"spread_pct":0.36490877587571585},{"cp_plan":{"max_rank_tokens":20916,"summary_ms":0.00036135315895080566,"wave_count":0},"fallback_work":775756288,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1336.16748046875,"n":8,"spread_pct":0.4439385617461321},{"cp_plan":{"max_rank_tokens":29108,"summary_ms":0.0005029141902923584,"wave_count":0},"fallback_work":1078336000,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1868.1707763671875,"n":8,"spread_pct":0.4792259306438718}],"cell":"cal-hetero|Qwen/Qwen3-8B|L36|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":36.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":12288,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":4096,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":32,"num_query_groups":8},"hidden_size":4096,"model":"Qwen/Qwen3-8B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}} + ], + "fit_arguments": { + "exclude_cells": ["cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g7","cal-grpo-g16|Qwen/Qwen3-8B|L36|tp2|cp2","cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp2|cp2","cal-grpo-g8|Qwen/Qwen3-8B|L36|tp2|cp2","cal-hetero2|Qwen/Qwen3-8B|L36|tp2|cp2","cal-hetero3|Qwen/Qwen3-8B|L36|tp2|cp2","cal-hetero|Qwen/Qwen3-8B|L36|tp2|cp2"], + "holdout": "", + "incumbent": "depth_one", + "objective": "regret", + "require_complete": 8, + "rerank_shapes": ["tp1cp4"], + "shortlist_size": 3, + "terms": ["token_per_rank","token_cp_exchange","token_tp_collective","gdn_token_per_rank","attention_token_cp_exchange","tiny_segment_per_layer","level_cp_per_layer","level_tp_per_layer","gdn_level","gdn_level_tp"] + }, + "integer_table_milli_us": {"attention_token_cp_exchange":0,"gdn_level":0,"gdn_level_tp":0,"gdn_token_per_rank":0,"level_cp_per_layer":0,"level_tp_per_layer":0,"tiny_segment_per_layer":0,"token_cp_exchange":262,"token_per_rank":2858,"token_tp_collective":640}, + "integer_table_sha256": "eb021d7e8165ba932ba75617e1ffc2831048eb3e8f689230578f8f7c73eef403", + "manifest": { + "excluded": ["cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g7","cal-grpo-g16|Qwen/Qwen3-8B|L36|tp2|cp2","cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp2|cp2","cal-grpo-g8|Qwen/Qwen3-8B|L36|tp2|cp2","cal-hetero2|Qwen/Qwen3-8B|L36|tp2|cp2","cal-hetero3|Qwen/Qwen3-8B|L36|tp2|cp2","cal-hetero|Qwen/Qwen3-8B|L36|tp2|cp2"], + "expected": ["cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g0","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g1","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g2","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g3","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g4","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g5","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g6","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g7","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g7","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g0","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g1","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g2","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g3","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g4","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g5","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g6","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g7","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g0","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g1","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g2","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g3","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g4","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g5","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g6","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g7","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g7","cal-grpo-g16|Qwen/Qwen3-8B|L36|tp1|cp1","cal-grpo-g16|Qwen/Qwen3-8B|L36|tp1|cp2","cal-grpo-g16|Qwen/Qwen3-8B|L36|tp1|cp4","cal-grpo-g16|Qwen/Qwen3-8B|L36|tp2|cp1","cal-grpo-g16|Qwen/Qwen3-8B|L36|tp2|cp2","cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp1|cp1","cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp1|cp2","cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp1|cp4","cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp2|cp1","cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp2|cp2","cal-grpo-g8|Qwen/Qwen3-8B|L36|tp1|cp1","cal-grpo-g8|Qwen/Qwen3-8B|L36|tp1|cp2","cal-grpo-g8|Qwen/Qwen3-8B|L36|tp1|cp4","cal-grpo-g8|Qwen/Qwen3-8B|L36|tp2|cp1","cal-grpo-g8|Qwen/Qwen3-8B|L36|tp2|cp2","cal-hetero2|Qwen/Qwen3-8B|L36|tp1|cp1","cal-hetero2|Qwen/Qwen3-8B|L36|tp1|cp2","cal-hetero2|Qwen/Qwen3-8B|L36|tp1|cp4","cal-hetero2|Qwen/Qwen3-8B|L36|tp2|cp1","cal-hetero2|Qwen/Qwen3-8B|L36|tp2|cp2","cal-hetero3|Qwen/Qwen3-8B|L36|tp1|cp1","cal-hetero3|Qwen/Qwen3-8B|L36|tp1|cp2","cal-hetero3|Qwen/Qwen3-8B|L36|tp1|cp4","cal-hetero3|Qwen/Qwen3-8B|L36|tp2|cp1","cal-hetero3|Qwen/Qwen3-8B|L36|tp2|cp2","cal-hetero|Qwen/Qwen3-8B|L36|tp1|cp1","cal-hetero|Qwen/Qwen3-8B|L36|tp1|cp2","cal-hetero|Qwen/Qwen3-8B|L36|tp1|cp4","cal-hetero|Qwen/Qwen3-8B|L36|tp2|cp1","cal-hetero|Qwen/Qwen3-8B|L36|tp2|cp2"], + "path": "trainer_rank_cost_calibration_manifest_dense-attn-h4096-h200-bf16.json" + }, + "measured_envelope": {"device_names":["NVIDIA H200"],"hidden_sizes":[4096],"param_dtypes":["torch.bfloat16"]}, + "metrics": { + "integer_all": {"cells":42,"clear_misses":[],"max_regret_pct":1.1113450032649341,"median_regret_pct":0.0,"ordered_pairs":3389,"p95_regret_pct":0.5378688884101532,"pairwise_accuracy":0.9994098554145766}, + "test": {"cells":12,"clear_misses":[],"max_regret_pct":1.1113450032649341,"median_regret_pct":0.0,"ordered_pairs":667,"p95_regret_pct":1.1113450032649341,"pairwise_accuracy":0.9985007496251874}, + "train": {"cells":30,"clear_misses":[],"max_regret_pct":0.9679103474105565,"median_regret_pct":0.0,"ordered_pairs":2722,"p95_regret_pct":0.5378688884101532,"pairwise_accuracy":0.9996326230712711} + }, + "reranker": {"incumbent":"depth_one","max_rank_token_per_layer_milli_us":2820,"metrics":{"all":{"cells":14,"clear_misses":[],"max_regret_pct":0.7545674588797711,"median_regret_pct":0.0,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":1.8360047494208644,"savings_vs_fallback_mean_pct":1.7865917180166857,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":33,"p95_regret_pct":0.7545674588797711,"pairwise_accuracy":0.9696969696969697,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g0":{"best":"minimum_effective_span_46","best_lead_pct":0.13999717453958127,"cheap_regret_pct":0.13999717453958127,"fallback_regret_pct":0.13999717453958127,"irrecoverable_pct":0.0,"planning_ms":8.20615328848362,"regret_pct":0.13999717453958127,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_46"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g1":{"best":"depth_one","best_lead_pct":44.849625027425496,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.811984211206436,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g2":{"best":"no_sharing","best_lead_pct":0.0,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":0.0,"regret_pct":0.0,"selected":"no_sharing","shortlist":["no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g3":{"best":"minimum_effective_span_278","best_lead_pct":5.955033521950754,"cheap_regret_pct":13.99212916951616,"fallback_regret_pct":13.99212916951616,"irrecoverable_pct":0.0,"planning_ms":22.93180488049984,"regret_pct":0.0,"selected":"minimum_effective_span_278","shortlist":["uniform_depth_3","minimum_effective_span_278","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g4":{"best":"uniform_depth_2","best_lead_pct":18.140498334185455,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.770554050803185,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_6717"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g5":{"best":"uniform_depth_3","best_lead_pct":0.5217273265133854,"cheap_regret_pct":0.7545674588797711,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":14.726538211107254,"regret_pct":0.7545674588797711,"selected":"minimum_effective_span_22","shortlist":["minimum_effective_span_22","uniform_depth_3","uniform_depth_4","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g6":{"best":"minimum_effective_span_71","best_lead_pct":0.20920858933146402,"cheap_regret_pct":11.711937322375942,"fallback_regret_pct":11.96439760143075,"irrecoverable_pct":0.0,"planning_ms":11.308472603559494,"regret_pct":0.0,"selected":"minimum_effective_span_71","shortlist":["minimum_effective_span_40","minimum_effective_span_71","minimum_effective_span_82","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g7":{"best":"minimum_effective_span_134","best_lead_pct":0.034609323744468706,"cheap_regret_pct":0.2885837104200407,"fallback_regret_pct":0.09890845058650127,"irrecoverable_pct":0.05284718818900615,"planning_ms":13.162782415747643,"regret_pct":0.2885837104200407,"selected":"minimum_effective_span_37","shortlist":["minimum_effective_span_37","minimum_effective_span_87","minimum_effective_span_53","depth_one"]},"cal-grpo-g16|Qwen/Qwen3-8B|L36|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":13.261221504332156,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.398846447467804,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":10.185789827095796,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.046437710523605,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g8|Qwen/Qwen3-8B|L36|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":18.2323543666529,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":9.922048076987267,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-hetero2|Qwen/Qwen3-8B|L36|tp1|cp4":{"best":"depth_one","best_lead_pct":3.5975916941142287,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":24.66788701713085,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_257","minimum_effective_span_1025"]},"cal-hetero3|Qwen/Qwen3-8B|L36|tp1|cp4":{"best":"depth_one","best_lead_pct":5.665019439206261,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":19.741032272577286,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_769","no_sharing"]},"cal-hetero|Qwen/Qwen3-8B|L36|tp1|cp4":{"best":"minimum_effective_span_513","best_lead_pct":0.09557940828990752,"cheap_regret_pct":0.09557940828990752,"fallback_regret_pct":0.09557940828990752,"irrecoverable_pct":0.0,"planning_ms":8.744480088353157,"regret_pct":0.09557940828990752,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_513","minimum_effective_span_2049"]}},"planning_cost":{"max_cell":"cal-hetero3|Qwen/Qwen3-8B|L36|tp1|cp4","max_ms":19.741032272577286,"max_pct_of_cell":2.626984893554178,"mean_pct_of_cell":1.5726608844450682,"median_pct_of_cell":1.635015289234426},"recall":{"best_in_shortlist":13,"cells":14,"clear_winner_cells":4,"clear_winners_in_shortlist":4,"max_irrecoverable_pct":0.05284718818900615}},"by_shape":{"tp1cp4":{"cells":14,"clear_misses":[],"max_regret_pct":0.7545674588797711,"median_regret_pct":0.0,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":1.8360047494208644,"savings_vs_fallback_mean_pct":1.7865917180166857,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":33,"p95_regret_pct":0.7545674588797711,"pairwise_accuracy":0.9696969696969697,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g0":{"best":"minimum_effective_span_46","best_lead_pct":0.13999717453958127,"cheap_regret_pct":0.13999717453958127,"fallback_regret_pct":0.13999717453958127,"irrecoverable_pct":0.0,"planning_ms":8.20615328848362,"regret_pct":0.13999717453958127,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_46"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g1":{"best":"depth_one","best_lead_pct":44.849625027425496,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.811984211206436,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g2":{"best":"no_sharing","best_lead_pct":0.0,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":0.0,"regret_pct":0.0,"selected":"no_sharing","shortlist":["no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g3":{"best":"minimum_effective_span_278","best_lead_pct":5.955033521950754,"cheap_regret_pct":13.99212916951616,"fallback_regret_pct":13.99212916951616,"irrecoverable_pct":0.0,"planning_ms":22.93180488049984,"regret_pct":0.0,"selected":"minimum_effective_span_278","shortlist":["uniform_depth_3","minimum_effective_span_278","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g4":{"best":"uniform_depth_2","best_lead_pct":18.140498334185455,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.770554050803185,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_6717"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g5":{"best":"uniform_depth_3","best_lead_pct":0.5217273265133854,"cheap_regret_pct":0.7545674588797711,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":14.726538211107254,"regret_pct":0.7545674588797711,"selected":"minimum_effective_span_22","shortlist":["minimum_effective_span_22","uniform_depth_3","uniform_depth_4","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g6":{"best":"minimum_effective_span_71","best_lead_pct":0.20920858933146402,"cheap_regret_pct":11.711937322375942,"fallback_regret_pct":11.96439760143075,"irrecoverable_pct":0.0,"planning_ms":11.308472603559494,"regret_pct":0.0,"selected":"minimum_effective_span_71","shortlist":["minimum_effective_span_40","minimum_effective_span_71","minimum_effective_span_82","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g7":{"best":"minimum_effective_span_134","best_lead_pct":0.034609323744468706,"cheap_regret_pct":0.2885837104200407,"fallback_regret_pct":0.09890845058650127,"irrecoverable_pct":0.05284718818900615,"planning_ms":13.162782415747643,"regret_pct":0.2885837104200407,"selected":"minimum_effective_span_37","shortlist":["minimum_effective_span_37","minimum_effective_span_87","minimum_effective_span_53","depth_one"]},"cal-grpo-g16|Qwen/Qwen3-8B|L36|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":13.261221504332156,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.398846447467804,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":10.185789827095796,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.046437710523605,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g8|Qwen/Qwen3-8B|L36|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":18.2323543666529,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":9.922048076987267,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-hetero2|Qwen/Qwen3-8B|L36|tp1|cp4":{"best":"depth_one","best_lead_pct":3.5975916941142287,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":24.66788701713085,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_257","minimum_effective_span_1025"]},"cal-hetero3|Qwen/Qwen3-8B|L36|tp1|cp4":{"best":"depth_one","best_lead_pct":5.665019439206261,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":19.741032272577286,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_769","no_sharing"]},"cal-hetero|Qwen/Qwen3-8B|L36|tp1|cp4":{"best":"minimum_effective_span_513","best_lead_pct":0.09557940828990752,"cheap_regret_pct":0.09557940828990752,"fallback_regret_pct":0.09557940828990752,"irrecoverable_pct":0.0,"planning_ms":8.744480088353157,"regret_pct":0.09557940828990752,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_513","minimum_effective_span_2049"]}},"planning_cost":{"max_cell":"cal-hetero3|Qwen/Qwen3-8B|L36|tp1|cp4","max_ms":19.741032272577286,"max_pct_of_cell":2.626984893554178,"mean_pct_of_cell":1.5726608844450682,"median_pct_of_cell":1.635015289234426},"recall":{"best_in_shortlist":13,"cells":14,"clear_winner_cells":4,"clear_winners_in_shortlist":4,"max_irrecoverable_pct":0.05284718818900615}}},"held_out":{"cells":4,"clear_misses":[],"max_regret_pct":0.7545674588797711,"median_regret_pct":0.14429185521002036,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":3.49803229237904,"savings_vs_fallback_mean_pct":3.2619716127007123,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":10,"p95_regret_pct":0.7545674588797711,"pairwise_accuracy":1.0,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g1":{"best":"depth_one","best_lead_pct":44.849625027425496,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.811984211206436,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g3":{"best":"minimum_effective_span_278","best_lead_pct":5.955033521950754,"cheap_regret_pct":13.99212916951616,"fallback_regret_pct":13.99212916951616,"irrecoverable_pct":0.0,"planning_ms":22.93180488049984,"regret_pct":0.0,"selected":"minimum_effective_span_278","shortlist":["uniform_depth_3","minimum_effective_span_278","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g5":{"best":"uniform_depth_3","best_lead_pct":0.5217273265133854,"cheap_regret_pct":0.7545674588797711,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":14.726538211107254,"regret_pct":0.7545674588797711,"selected":"minimum_effective_span_22","shortlist":["minimum_effective_span_22","uniform_depth_3","uniform_depth_4","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g7":{"best":"minimum_effective_span_134","best_lead_pct":0.034609323744468706,"cheap_regret_pct":0.2885837104200407,"fallback_regret_pct":0.09890845058650127,"irrecoverable_pct":0.05284718818900615,"planning_ms":13.162782415747643,"regret_pct":0.2885837104200407,"selected":"minimum_effective_span_37","shortlist":["minimum_effective_span_37","minimum_effective_span_87","minimum_effective_span_53","depth_one"]}},"planning_cost":{"max_cell":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g3","max_ms":22.93180488049984,"max_pct_of_cell":2.5449986612496693,"mean_pct_of_cell":1.756666437875416,"median_pct_of_cell":1.7995891626245364},"recall":{"best_in_shortlist":3,"cells":4,"clear_winner_cells":2,"clear_winners_in_shortlist":2,"max_irrecoverable_pct":0.05284718818900615}}},"shapes":["tp1cp4"],"shortlist_size":3,"shortlist_table_milli_us":{"attention_token_cp_exchange":57,"gdn_level":0,"gdn_level_tp":0,"gdn_token_per_rank":0,"level_cp_per_layer":17614,"level_tp_per_layer":0,"tiny_segment_per_layer":1508,"token_cp_exchange":0,"token_per_rank":2925,"token_tp_collective":573},"wave_per_layer_milli_us":249653}, + "schema": "art.dev.trainer_rank_cost_calibration_certificate.v1", + "table_id": "dense-attn-h4096-h200-bf16" +} diff --git a/dev/trainer_rank_cost_calibration_certificate_dense-attn-h5120-h200-bf16.json b/dev/trainer_rank_cost_calibration_certificate_dense-attn-h5120-h200-bf16.json new file mode 100644 index 000000000..f1083c90b --- /dev/null +++ b/dev/trainer_rank_cost_calibration_certificate_dense-attn-h5120-h200-bf16.json @@ -0,0 +1,101 @@ +{ + "admitted": {"device_classes":[{"capability":[9,0],"memory_class":"hbm-141g"}],"geometries":[{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8}],"param_dtypes":["torch.bfloat16"],"shapes":[[1,1,1,1],[1,2,1,1],[1,4,1,1],[2,1,1,1],[2,2,1,1]]}, + "cells": [ + {"candidates":[{"cp_plan":{"max_rank_tokens":12204,"summary_ms":0.00042654573917388916,"wave_count":0},"fallback_work":502191616,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":2510.02001953125,"n":8,"spread_pct":0.44364180206636644},{"cp_plan":{"max_rank_tokens":23926,"summary_ms":0.00036694109439849854,"wave_count":0},"fallback_work":983825152,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":4849.888671875,"n":8,"spread_pct":0.27856833053915947},{"cp_plan":{"max_rank_tokens":35778,"summary_ms":0.00042654573917388916,"wave_count":0},"fallback_work":1470800128,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":7216.350830078125,"n":8,"spread_pct":0.24067793364630488},{"cp_plan":{"max_rank_tokens":12154,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":500366592,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":2489.312744140625,"n":8,"spread_pct":0.4784319686942962},{"cp_plan":{"max_rank_tokens":47640,"summary_ms":0.0013709068298339844,"wave_count":0},"fallback_work":1957956608,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":9784.98388671875,"n":8,"spread_pct":0.43565637658188294},{"cp_plan":{"max_rank_tokens":12114,"summary_ms":0.0005867332220077515,"wave_count":0},"fallback_work":498723072,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":2484.2598876953125,"n":8,"spread_pct":0.3689439816320106},{"cp_plan":{"max_rank_tokens":12109,"summary_ms":0.00037997961044311523,"wave_count":0},"fallback_work":498747008,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":2482.6612548828125,"n":8,"spread_pct":0.25053642299736933}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11920,"summary_ms":0.00037997961044311523,"wave_count":0},"fallback_work":490260480,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":2420.520751953125,"n":8,"spread_pct":1.4025887257723648},{"cp_plan":{"max_rank_tokens":23728,"summary_ms":0.0010263174772262573,"wave_count":0},"fallback_work":975198208,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":4802.29296875,"n":8,"spread_pct":0.8436826646125264}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12138,"summary_ms":0.0006966292858123779,"wave_count":0},"fallback_work":498857216,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2491.2254638671875,"n":8,"spread_pct":0.7087081371879496}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":20701,"summary_ms":0.0003296881914138794,"wave_count":0},"fallback_work":851447424,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":4307.4296875,"n":8,"spread_pct":0.16071885241887654},{"cp_plan":{"max_rank_tokens":16154,"summary_ms":0.00025331974029541016,"wave_count":0},"fallback_work":664849664,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":3333.1314697265625,"n":8,"spread_pct":0.23427199979725336},{"cp_plan":{"max_rank_tokens":14771,"summary_ms":0.0002216547727584839,"wave_count":0},"fallback_work":608254336,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":3026.33349609375,"n":8,"spread_pct":0.10456715313545761},{"cp_plan":{"max_rank_tokens":21963,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":903529856,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":4371.872314453125,"n":8,"spread_pct":0.3915075671564085},{"cp_plan":{"max_rank_tokens":27497,"summary_ms":0.00018440186977386475,"wave_count":0},"fallback_work":1130910848,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":5377.57373046875,"n":8,"spread_pct":0.38088593298031076},{"cp_plan":{"max_rank_tokens":35243,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":1448949120,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":6878.775634765625,"n":8,"spread_pct":0.2046745279980295},{"cp_plan":{"max_rank_tokens":46153,"summary_ms":0.0010710209608078003,"wave_count":0},"fallback_work":1896989824,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":9172.3935546875,"n":8,"spread_pct":0.1419851582070903},{"cp_plan":{"max_rank_tokens":15600,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":642316288,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":3192.6986083984375,"n":8,"spread_pct":0.349774209778193},{"cp_plan":{"max_rank_tokens":14494,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":597102336,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":2983.6671142578125,"n":8,"spread_pct":0.22794182897031245}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16800,"summary_ms":0.0003203749656677246,"wave_count":0},"fallback_work":690900992,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":3487.15966796875,"n":8,"spread_pct":0.12590834427026718},{"cp_plan":{"max_rank_tokens":18971,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":780103040,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":3743.0338134765625,"n":8,"spread_pct":0.3154949332472533},{"cp_plan":{"max_rank_tokens":30232,"summary_ms":0.0006593763828277588,"wave_count":0},"fallback_work":1242565632,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":6010.623291015625,"n":8,"spread_pct":0.4906189655335587},{"cp_plan":{"max_rank_tokens":12255,"summary_ms":0.00019744038581848145,"wave_count":0},"fallback_work":504385408,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":2488.527587890625,"n":8,"spread_pct":0.2889333267536603}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13078,"summary_ms":0.00034086406230926514,"wave_count":0},"fallback_work":538233600,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":2718.6187744140625,"n":8,"spread_pct":0.32992792171856355},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":1028577792,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":5093.605224609375,"n":8,"spread_pct":0.15821001814324956},{"cp_plan":{"max_rank_tokens":36716,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":1509471744,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":7404.089111328125,"n":8,"spread_pct":0.17212959912585254},{"cp_plan":{"max_rank_tokens":48890,"summary_ms":0.00027567148208618164,"wave_count":0},"fallback_work":2009677056,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":9942.34033203125,"n":8,"spread_pct":0.2762412352906134},{"cp_plan":{"max_rank_tokens":12899,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":531108224,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":2677.689453125,"n":8,"spread_pct":0.3247136895804963},{"cp_plan":{"max_rank_tokens":12722,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":523835648,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":2617.9208984375,"n":8,"spread_pct":0.4094095274048205},{"cp_plan":{"max_rank_tokens":61066,"summary_ms":0.0008046627044677734,"wave_count":0},"fallback_work":2509735168,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":12575.1171875,"n":8,"spread_pct":0.09324434725074009},{"cp_plan":{"max_rank_tokens":13015,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":535874432,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":2732.843994140625,"n":8,"spread_pct":0.37744347750423474},{"cp_plan":{"max_rank_tokens":12701,"summary_ms":0.00027008354663848877,"wave_count":0},"fallback_work":523202176,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":2617.478515625,"n":8,"spread_pct":0.374324049042308},{"cp_plan":{"max_rank_tokens":12700,"summary_ms":0.00028312206268310547,"wave_count":0},"fallback_work":523390464,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":2618.6329345703125,"n":8,"spread_pct":0.3805267808949358}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9058,"summary_ms":0.0005308538675308228,"wave_count":0},"fallback_work":374108416,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1772.9755859375,"n":8,"spread_pct":0.9492078608580319},{"cp_plan":{"max_rank_tokens":6403,"summary_ms":0.00027567148208618164,"wave_count":0},"fallback_work":265937280,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":1271.62060546875,"n":8,"spread_pct":1.5322091066279764},{"cp_plan":{"max_rank_tokens":6443,"summary_ms":0.0002551823854446411,"wave_count":0},"fallback_work":267580800,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":1274.5725708007812,"n":8,"spread_pct":1.7055541511431744},{"cp_plan":{"max_rank_tokens":6668,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":276596224,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":1307.4915771484375,"n":8,"spread_pct":1.40407438317217},{"cp_plan":{"max_rank_tokens":6755,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":280170880,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":1329.091796875,"n":8,"spread_pct":0.8658219384418695},{"cp_plan":{"max_rank_tokens":6799,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":281978752,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":1336.4878540039062,"n":8,"spread_pct":1.1025154704862312},{"cp_plan":{"max_rank_tokens":7040,"summary_ms":0.00029616057872772217,"wave_count":0},"fallback_work":291651584,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":1402.0861206054688,"n":8,"spread_pct":0.9701468919755843},{"cp_plan":{"max_rank_tokens":7291,"summary_ms":0.00026263296604156494,"wave_count":0},"fallback_work":301964672,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":1415.9845581054688,"n":8,"spread_pct":0.8544325390587233},{"cp_plan":{"max_rank_tokens":7367,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":305087360,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":1451.9657592773438,"n":8,"spread_pct":1.4347131008873133},{"cp_plan":{"max_rank_tokens":7666,"summary_ms":0.00026263296604156494,"wave_count":0},"fallback_work":317143296,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":1504.9132080078125,"n":8,"spread_pct":0.9043782292197375},{"cp_plan":{"max_rank_tokens":5988,"summary_ms":0.00026263296604156494,"wave_count":0},"fallback_work":250262016,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":1181.8230590820312,"n":8,"spread_pct":1.455901255741043},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.0002551823854446411,"wave_count":0},"fallback_work":320101632,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":1523.9723510742188,"n":8,"spread_pct":1.1187738855945726},{"cp_plan":{"max_rank_tokens":6024,"summary_ms":0.00025704503059387207,"wave_count":0},"fallback_work":251511808,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":1181.9710083007812,"n":8,"spread_pct":0.9794988622241105},{"cp_plan":{"max_rank_tokens":6102,"summary_ms":0.0002551823854446411,"wave_count":0},"fallback_work":254257920,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":1214.0198364257812,"n":8,"spread_pct":1.6034286919168101},{"cp_plan":{"max_rank_tokens":8016,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":331524096,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":1571.91748046875,"n":8,"spread_pct":0.5915983880902821},{"cp_plan":{"max_rank_tokens":13361,"summary_ms":0.000303611159324646,"wave_count":0},"fallback_work":550910080,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":2517.4373779296875,"n":8,"spread_pct":0.7214910724983801},{"cp_plan":{"max_rank_tokens":17592,"summary_ms":0.0003855675458908081,"wave_count":0},"fallback_work":724753408,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":3264.8990478515625,"n":8,"spread_pct":0.8838238673945648},{"cp_plan":{"max_rank_tokens":22066,"summary_ms":0.0003241002559661865,"wave_count":0},"fallback_work":908581120,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":4057.20263671875,"n":8,"spread_pct":0.7855736404357075},{"cp_plan":{"max_rank_tokens":26682,"summary_ms":0.0004377216100692749,"wave_count":0},"fallback_work":1098243328,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":4901.52587890625,"n":8,"spread_pct":0.9396388341680508},{"cp_plan":{"max_rank_tokens":31360,"summary_ms":0.0003241002559661865,"wave_count":0},"fallback_work":1290452992,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":5850.9130859375,"n":8,"spread_pct":0.6640590118896706},{"cp_plan":{"max_rank_tokens":35986,"summary_ms":0.00027567148208618164,"wave_count":0},"fallback_work":1480526080,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":6625.514404296875,"n":8,"spread_pct":0.43731863832654205},{"cp_plan":{"max_rank_tokens":40782,"summary_ms":0.0002998858690261841,"wave_count":0},"fallback_work":1677584128,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":7443.524169921875,"n":8,"spread_pct":0.7915135702027063},{"cp_plan":{"max_rank_tokens":45522,"summary_ms":0.0003427267074584961,"wave_count":0},"fallback_work":1872341248,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":8353.31494140625,"n":8,"spread_pct":0.6649095907085446},{"cp_plan":{"max_rank_tokens":50326,"summary_ms":0.00027567148208618164,"wave_count":0},"fallback_work":2069728000,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":9179.77197265625,"n":8,"spread_pct":0.8363114530914157},{"cp_plan":{"max_rank_tokens":55336,"summary_ms":0.0002514570951461792,"wave_count":0},"fallback_work":2275578880,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":10032.87548828125,"n":8,"spread_pct":0.6478408945513064},{"cp_plan":{"max_rank_tokens":8411,"summary_ms":0.00025704503059387207,"wave_count":0},"fallback_work":347753856,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":1660.7630615234375,"n":8,"spread_pct":0.594069623522167},{"cp_plan":{"max_rank_tokens":60463,"summary_ms":0.0002980232238769531,"wave_count":0},"fallback_work":2486237056,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":10968.16357421875,"n":8,"spread_pct":0.7897947179245116},{"cp_plan":{"max_rank_tokens":6223,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":259000192,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":1231.0059814453125,"n":8,"spread_pct":1.4845006659294886},{"cp_plan":{"max_rank_tokens":6304,"summary_ms":0.000260770320892334,"wave_count":0},"fallback_work":262098944,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":1265.0763549804688,"n":8,"spread_pct":1.9796583308611164},{"cp_plan":{"max_rank_tokens":6322,"summary_ms":0.00030919909477233887,"wave_count":0},"fallback_work":262838528,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":1254.042236328125,"n":8,"spread_pct":0.5040535907612823},{"cp_plan":{"max_rank_tokens":6333,"summary_ms":0.0002551823854446411,"wave_count":0},"fallback_work":263290496,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":1257.9255981445312,"n":8,"spread_pct":1.5053770640673663},{"cp_plan":{"max_rank_tokens":6310,"summary_ms":0.00025890767574310303,"wave_count":0},"fallback_work":262345472,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":1255.7561645507812,"n":8,"spread_pct":1.5943670331771531},{"cp_plan":{"max_rank_tokens":65842,"summary_ms":0.001169741153717041,"wave_count":0},"fallback_work":2707020032,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":12233.5771484375,"n":8,"spread_pct":0.5270298996989187},{"cp_plan":{"max_rank_tokens":6245,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":260592256,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":1237.6181640625,"n":8,"spread_pct":1.3637626142074502},{"cp_plan":{"max_rank_tokens":6083,"summary_ms":0.0002514570951461792,"wave_count":0},"fallback_work":254165376,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":1211.78173828125,"n":8,"spread_pct":1.232245744801306},{"cp_plan":{"max_rank_tokens":5957,"summary_ms":0.00028312206268310547,"wave_count":0},"fallback_work":249217664,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":1178.17431640625,"n":8,"spread_pct":0.6182081237481356},{"cp_plan":{"max_rank_tokens":8629,"summary_ms":0.00028870999813079834,"wave_count":0},"fallback_work":356711040,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1686.5258178710938,"n":8,"spread_pct":1.2612026206245583},{"cp_plan":{"max_rank_tokens":8089,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":334752896,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1585.2471313476562,"n":8,"spread_pct":0.7285113990314981},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":320560384,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":1523.073486328125,"n":8,"spread_pct":1.5481263231326587},{"cp_plan":{"max_rank_tokens":7490,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":310599936,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":1476.6166381835938,"n":8,"spread_pct":1.1167414620669291},{"cp_plan":{"max_rank_tokens":7273,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":301913216,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":1418.7445678710938,"n":8,"spread_pct":0.6225588263928082},{"cp_plan":{"max_rank_tokens":6973,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":289816192,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":1360.7941284179688,"n":8,"spread_pct":0.6761260503413934},{"cp_plan":{"max_rank_tokens":6793,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":282649728,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":1332.1209716796875,"n":8,"spread_pct":0.36962319711786273},{"cp_plan":{"max_rank_tokens":6521,"summary_ms":0.00025704503059387207,"wave_count":0},"fallback_work":271703168,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":1288.7841796875,"n":8,"spread_pct":0.9494095316034145}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12291,"summary_ms":0.0003762543201446533,"wave_count":0},"fallback_work":506290560,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":2548.4241943359375,"n":8,"spread_pct":0.7510481426390802},{"cp_plan":{"max_rank_tokens":23563,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":969434496,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":4791.885009765625,"n":8,"spread_pct":0.8305463934107757},{"cp_plan":{"max_rank_tokens":34902,"summary_ms":0.00025331974029541016,"wave_count":0},"fallback_work":1435331328,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":7065.712646484375,"n":8,"spread_pct":0.6743268298273749},{"cp_plan":{"max_rank_tokens":46161,"summary_ms":0.000476837158203125,"wave_count":0},"fallback_work":1897941120,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":9302.6943359375,"n":8,"spread_pct":0.6447224893577643},{"cp_plan":{"max_rank_tokens":57385,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":2359112832,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":11552.6533203125,"n":8,"spread_pct":0.7033692628612694},{"cp_plan":{"max_rank_tokens":68715,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":2824639872,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":13897.671875,"n":8,"spread_pct":0.6057950353645114},{"cp_plan":{"max_rank_tokens":80181,"summary_ms":0.0002477318048477173,"wave_count":0},"fallback_work":3295754880,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":16231.71875,"n":8,"spread_pct":0.6697618955940817},{"cp_plan":{"max_rank_tokens":11890,"summary_ms":0.00028312206268310547,"wave_count":0},"fallback_work":490731776,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":2472.5938720703125,"n":8,"spread_pct":1.0258355154990333},{"cp_plan":{"max_rank_tokens":12153,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":500849792,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":2506.11181640625,"n":8,"spread_pct":0.9212244206887186},{"cp_plan":{"max_rank_tokens":11903,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":491036544,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":2473.4593505859375,"n":8,"spread_pct":0.4352259403899178},{"cp_plan":{"max_rank_tokens":11876,"summary_ms":0.0002644956111907959,"wave_count":0},"fallback_work":490385920,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":2458.3817138671875,"n":8,"spread_pct":1.0440803526956544},{"cp_plan":{"max_rank_tokens":11939,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":492286336,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":2472.447998046875,"n":8,"spread_pct":1.0060286475549332},{"cp_plan":{"max_rank_tokens":11981,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":494012032,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":2481.213623046875,"n":8,"spread_pct":0.9816343536994944},{"cp_plan":{"max_rank_tokens":12033,"summary_ms":0.00024586915969848633,"wave_count":0},"fallback_work":495919232,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":2490.835693359375,"n":8,"spread_pct":1.3565351817497402},{"cp_plan":{"max_rank_tokens":11881,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":490361984,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":2473.276123046875,"n":8,"spread_pct":0.8543770474601599},{"cp_plan":{"max_rank_tokens":12025,"summary_ms":0.00026263296604156494,"wave_count":0},"fallback_work":495590528,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":2489.0233154296875,"n":8,"spread_pct":1.1716874662427743},{"cp_plan":{"max_rank_tokens":91657,"summary_ms":0.0008717179298400879,"wave_count":0},"fallback_work":3767051392,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":18860.5673828125,"n":8,"spread_pct":0.6141181732451568},{"cp_plan":{"max_rank_tokens":12225,"summary_ms":0.0002980232238769531,"wave_count":0},"fallback_work":503808128,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":2543.7259521484375,"n":8,"spread_pct":1.0263744911209474},{"cp_plan":{"max_rank_tokens":12215,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":503626624,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":2539.3018798828125,"n":8,"spread_pct":1.0478531326926341},{"cp_plan":{"max_rank_tokens":12123,"summary_ms":0.000303611159324646,"wave_count":0},"fallback_work":500075904,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":2512.2860107421875,"n":8,"spread_pct":0.9696099447213686},{"cp_plan":{"max_rank_tokens":11973,"summary_ms":0.00025890767574310303,"wave_count":0},"fallback_work":494142080,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":2482.1834716796875,"n":8,"spread_pct":1.1713162657684426},{"cp_plan":{"max_rank_tokens":11879,"summary_ms":0.000260770320892334,"wave_count":0},"fallback_work":490509184,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":2472.0181884765625,"n":8,"spread_pct":0.8114356918622316},{"cp_plan":{"max_rank_tokens":11874,"summary_ms":0.00027008354663848877,"wave_count":0},"fallback_work":490533120,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":2470.6358642578125,"n":8,"spread_pct":0.8627211145916077}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.846645683050156,"wave_count":2},"fallback_work":501802432,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":2124.7974853515625,"n":8,"spread_pct":0.5978625835263138},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":12.316329404711723,"wave_count":1},"fallback_work":982873312,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":3330.4793701171875,"n":8,"spread_pct":0.8166098168450159},{"cp_plan":{"max_rank_tokens":20480,"summary_ms":5.142809823155403,"wave_count":1},"fallback_work":1469279392,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":4469.140380859375,"n":8,"spread_pct":0.6155384480999502},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.375146865844727,"wave_count":2},"fallback_work":500045344,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":2124.85302734375,"n":8,"spread_pct":0.9638764078122132},{"cp_plan":{"max_rank_tokens":27136,"summary_ms":8.560605347156525,"wave_count":1},"fallback_work":1955800960,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5890.36767578125,"n":8,"spread_pct":0.45376650654417683},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.428548902273178,"wave_count":2},"fallback_work":498403744,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":2120.6756591796875,"n":8,"spread_pct":0.7559846482170893},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.409101024270058,"wave_count":2},"fallback_work":498493536,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":2122.229736328125,"n":8,"spread_pct":0.7677748093706259}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.647787988185883,"wave_count":2},"fallback_work":489819392,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":2108.0946044921875,"n":8,"spread_pct":0.4854682914878134},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":9.818920865654945,"wave_count":1},"fallback_work":974124800,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3315.802734375,"n":8,"spread_pct":0.45500035047602894}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.181800574064255,"wave_count":2},"fallback_work":498307360,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2061.135498046875,"n":8,"spread_pct":0.6945646214096897}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13824,"summary_ms":8.657127618789673,"wave_count":1},"fallback_work":850683232,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":3089.4281005859375,"n":8,"spread_pct":0.9146932099549581},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.418081000447273,"wave_count":2},"fallback_work":664369184,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":2501.221435546875,"n":8,"spread_pct":1.2395994417961615},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":5.787491798400879,"wave_count":2},"fallback_work":607905856,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":2366.7479248046875,"n":8,"spread_pct":0.5557447007357338},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.18777796626091,"wave_count":1},"fallback_work":902770624,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":2488.419921875,"n":8,"spread_pct":0.8247087485705271},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":7.764901965856552,"wave_count":4},"fallback_work":1129885984,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":3017.781982421875,"n":8,"spread_pct":0.608510671002404},{"cp_plan":{"max_rank_tokens":17835,"summary_ms":5.393855273723602,"wave_count":1},"fallback_work":1447486912,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":3774.79296875,"n":8,"spread_pct":0.32846637905431486},{"cp_plan":{"max_rank_tokens":24649,"summary_ms":6.837707012891769,"wave_count":1},"fallback_work":1894938400,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":5220.594970703125,"n":8,"spread_pct":0.32500673297711885},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":7.242552936077118,"wave_count":3},"fallback_work":641927936,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":2368.34228515625,"n":8,"spread_pct":1.3035926162257894},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":7.182428613305092,"wave_count":3},"fallback_work":596832608,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":2262.023193359375,"n":8,"spread_pct":0.8312136740661096}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11776,"summary_ms":6.023179739713669,"wave_count":2},"fallback_work":690258432,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":2610.30712890625,"n":8,"spread_pct":0.3008188130400274},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":6.118046119809151,"wave_count":1},"fallback_work":779356352,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":2969.964599609375,"n":8,"spread_pct":0.6121426472310876},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":7.2176288813352585,"wave_count":1},"fallback_work":1241212800,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":3834.2254638671875,"n":8,"spread_pct":0.4467434310278821},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.695942789316177,"wave_count":2},"fallback_work":504026624,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":2125.9012451171875,"n":8,"spread_pct":0.5166352069317109}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.510639891028404,"wave_count":2},"fallback_work":537835232,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":2275.063720703125,"n":8,"spread_pct":0.756911320199329},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":11.572567746043205,"wave_count":1},"fallback_work":1027606592,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":2930.1861572265625,"n":8,"spread_pct":0.8193522060932545},{"cp_plan":{"max_rank_tokens":21504,"summary_ms":8.58324021100998,"wave_count":1},"fallback_work":1507938752,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":4724.487548828125,"n":8,"spread_pct":0.5367438939232672},{"cp_plan":{"max_rank_tokens":25850,"summary_ms":11.356085538864136,"wave_count":1},"fallback_work":2007559712,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":5792.384033203125,"n":8,"spread_pct":0.9527428278435892},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.464550599455833,"wave_count":2},"fallback_work":530784064,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":2273.165771484375,"n":8,"spread_pct":0.8337444112467481},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.243128657341003,"wave_count":2},"fallback_work":523519904,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":2274.406494140625,"n":8,"spread_pct":22.387397381850455},{"cp_plan":{"max_rank_tokens":33280,"summary_ms":7.430890575051308,"wave_count":1},"fallback_work":2506967840,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":7305.22216796875,"n":8,"spread_pct":0.7157965376197397},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.487457409501076,"wave_count":2},"fallback_work":535544704,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":2276.0771484375,"n":8,"spread_pct":0.6895880770737413},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.27130302786827,"wave_count":2},"fallback_work":522953056,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":2273.0125732421875,"n":8,"spread_pct":0.6134630296824701},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.355164900422096,"wave_count":2},"fallback_work":523206848,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":2273.4437255859375,"n":8,"spread_pct":0.7392269275894917}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7680,"summary_ms":9.96684655547142,"wave_count":1},"fallback_work":374165152,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1651.9591064453125,"n":8,"spread_pct":0.8706666275519662},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.969118162989616,"wave_count":2},"fallback_work":266383680,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":1338.7299194335938,"n":8,"spread_pct":0.8709408618708219},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.038043484091759,"wave_count":2},"fallback_work":268025280,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":1336.7942504882812,"n":8,"spread_pct":0.9592988999151083},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.184633821249008,"wave_count":2},"fallback_work":276964288,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":1338.0311889648438,"n":8,"spread_pct":1.6542148879157452},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.134862080216408,"wave_count":2},"fallback_work":280534848,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":1332.3118286132812,"n":8,"spread_pct":1.8715724148464512},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.127512082457542,"wave_count":2},"fallback_work":282340608,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":1337.426513671875,"n":8,"spread_pct":1.287882895143005},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.558459252119064,"wave_count":2},"fallback_work":291936256,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":1440.3388061523438,"n":8,"spread_pct":1.077576462345442},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.8237724006175995,"wave_count":2},"fallback_work":302237376,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":1430.8103637695312,"n":8,"spread_pct":41.789506422914},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.855008959770203,"wave_count":2},"fallback_work":305356416,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":1431.915283203125,"n":8,"spread_pct":1.3926386905649921},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.147487089037895,"wave_count":2},"fallback_work":317332384,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":1549.8491821289062,"n":8,"spread_pct":0.6970341785634777},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.596522241830826,"wave_count":1},"fallback_work":251121472,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":1229.870361328125,"n":8,"spread_pct":51.64460931298137},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":7.4815526604652405,"wave_count":1},"fallback_work":320287264,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":1541.1767578125,"n":8,"spread_pct":1.236428066773753},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.590252578258514,"wave_count":1},"fallback_work":252304000,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":1235.6395263671875,"n":8,"spread_pct":1.9752806633315416},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.070323124527931,"wave_count":1},"fallback_work":254915296,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":1220.894775390625,"n":8,"spread_pct":2.042269363709898},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.825817748904228,"wave_count":1},"fallback_work":331696384,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":1542.6144409179688,"n":8,"spread_pct":1.0587874312119536},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":5.7042669504880905,"wave_count":2},"fallback_work":550760352,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":2214.0556640625,"n":8,"spread_pct":1.0336790912815979},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":6.220193579792976,"wave_count":1},"fallback_work":724400512,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":2585.9063720703125,"n":8,"spread_pct":0.7768503633328291},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":8.358996361494064,"wave_count":1},"fallback_work":908013472,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":2765.0753173828125,"n":8,"spread_pct":1.0046400023804165},{"cp_plan":{"max_rank_tokens":16896,"summary_ms":7.345080375671387,"wave_count":1},"fallback_work":1097454112,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":3351.1893310546875,"n":8,"spread_pct":0.6930550214598122},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":8.273543789982796,"wave_count":1},"fallback_work":1289439232,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":3670.7840576171875,"n":8,"spread_pct":0.9989603626582805},{"cp_plan":{"max_rank_tokens":20992,"summary_ms":7.485495880246162,"wave_count":1},"fallback_work":1479290272,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":4169.508544921875,"n":8,"spread_pct":0.8042483367935709},{"cp_plan":{"max_rank_tokens":23040,"summary_ms":7.440589368343353,"wave_count":1},"fallback_work":1676118112,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":4582.994140625,"n":8,"spread_pct":0.5334237093398532},{"cp_plan":{"max_rank_tokens":26112,"summary_ms":7.117686793208122,"wave_count":2},"fallback_work":1870647712,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":5193.341796875,"n":8,"spread_pct":0.7065932152459335},{"cp_plan":{"max_rank_tokens":27648,"summary_ms":8.347522467374802,"wave_count":1},"fallback_work":2067803872,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":5499.646484375,"n":8,"spread_pct":0.9292855826315541},{"cp_plan":{"max_rank_tokens":30720,"summary_ms":10.730205103754997,"wave_count":1},"fallback_work":2273414272,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":6081.5634765625,"n":8,"spread_pct":0.9029676495671392},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.309119984507561,"wave_count":1},"fallback_work":347907264,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":1536.6127319335938,"n":8,"spread_pct":0.9991395329725141},{"cp_plan":{"max_rank_tokens":32768,"summary_ms":9.342215955257416,"wave_count":1},"fallback_work":2483826432,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":6458.2041015625,"n":8,"spread_pct":0.6456101969379126},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.099157199263573,"wave_count":1},"fallback_work":259586304,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":1332.7353515625,"n":8,"spread_pct":1.1413779812870177},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.382280796766281,"wave_count":1},"fallback_work":262615552,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":1329.7684936523438,"n":8,"spread_pct":0.6731467911927185},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.034374073147774,"wave_count":1},"fallback_work":263354272,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":1328.5819702148438,"n":8,"spread_pct":1.642403645400133},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.0634946674108505,"wave_count":1},"fallback_work":263805792,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":1332.7493286132812,"n":8,"spread_pct":1.0991600248491409},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.352219566702843,"wave_count":1},"fallback_work":262861792,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":1332.0584106445312,"n":8,"spread_pct":0.792304225834838},{"cp_plan":{"max_rank_tokens":35328,"summary_ms":8.265148848295212,"wave_count":1},"fallback_work":2704285600,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":7031.584228515625,"n":8,"spread_pct":0.7226114841971575},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.353944703936577,"wave_count":1},"fallback_work":261373920,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":1332.1173706054688,"n":8,"spread_pct":1.5394166002188916},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.291499197483063,"wave_count":1},"fallback_work":255020352,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":1224.829345703125,"n":8,"spread_pct":0.7942652246875113},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.703132599592209,"wave_count":1},"fallback_work":250144224,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":1232.9317016601562,"n":8,"spread_pct":1.8188989073606783},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":9.051617234945297,"wave_count":1},"fallback_work":356853984,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1539.76123046875,"n":8,"spread_pct":0.7297606955725366},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":9.895944967865944,"wave_count":1},"fallback_work":334987296,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1543.8316040039062,"n":8,"spread_pct":0.7744482446663431},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":9.759731590747833,"wave_count":1},"fallback_work":320877088,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":1428.8788452148438,"n":8,"spread_pct":1.4226198894687316},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":9.319685399532318,"wave_count":2},"fallback_work":310994080,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":1433.865478515625,"n":8,"spread_pct":1.180259940511957},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":9.802870452404022,"wave_count":2},"fallback_work":302383392,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":1435.6550903320312,"n":8,"spread_pct":0.9678267025194062},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":10.147171095013618,"wave_count":2},"fallback_work":290366304,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":1439.2883911132812,"n":8,"spread_pct":40.75368869597144},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":9.968696162104607,"wave_count":2},"fallback_work":283274016,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":1337.0174560546875,"n":8,"spread_pct":2.0892924004142484},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.110651582479477,"wave_count":2},"fallback_work":272406048,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":1338.6822509765625,"n":8,"spread_pct":0.5997272599196553}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.063295528292656,"wave_count":2},"fallback_work":506028352,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":2147.549072265625,"n":8,"spread_pct":0.8806937323286299},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":9.366348385810852,"wave_count":1},"fallback_work":968631232,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":3319.001220703125,"n":8,"spread_pct":0.3640996749525125},{"cp_plan":{"max_rank_tokens":20480,"summary_ms":6.592603400349617,"wave_count":1},"fallback_work":1433983712,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":4457.578857421875,"n":8,"spread_pct":0.2987253483262327},{"cp_plan":{"max_rank_tokens":26624,"summary_ms":9.586259722709656,"wave_count":1},"fallback_work":1896053152,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":5797.42724609375,"n":8,"spread_pct":0.6627822096799909},{"cp_plan":{"max_rank_tokens":31232,"summary_ms":7.010079920291901,"wave_count":1},"fallback_work":2356686112,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":6758.80712890625,"n":8,"spread_pct":0.28318086420372957},{"cp_plan":{"max_rank_tokens":36864,"summary_ms":11.836260557174683,"wave_count":1},"fallback_work":2821669312,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":7999.6826171875,"n":8,"spread_pct":0.15147085303066835},{"cp_plan":{"max_rank_tokens":43008,"summary_ms":9.141962975263596,"wave_count":1},"fallback_work":3292233952,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":9349.78662109375,"n":8,"spread_pct":6.146081521433451},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.3236672431230545,"wave_count":2},"fallback_work":490750880,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":2028.4194946289062,"n":8,"spread_pct":0.7060559899800278},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.417428910732269,"wave_count":2},"fallback_work":500659744,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":2151.39208984375,"n":8,"spread_pct":0.6942044020824998},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.202671676874161,"wave_count":2},"fallback_work":490989568,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":2030.41552734375,"n":8,"spread_pct":0.38073934700577716},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.302151829004288,"wave_count":2},"fallback_work":490471232,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":2029.5408325195312,"n":8,"spread_pct":28.87013973927641},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.531188264489174,"wave_count":2},"fallback_work":492172096,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":2146.427490234375,"n":8,"spread_pct":0.5029250661442689},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.524832755327225,"wave_count":2},"fallback_work":493895776,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":2147.574462890625,"n":8,"spread_pct":0.81150780455558},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.453102290630341,"wave_count":2},"fallback_work":495734944,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":2147.2467041015625,"n":8,"spread_pct":0.4282600952971377},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.357328966259956,"wave_count":2},"fallback_work":490381600,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":2029.6891479492188,"n":8,"spread_pct":1.0101031059714245},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.448341369628906,"wave_count":2},"fallback_work":495406624,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":2149.696044921875,"n":8,"spread_pct":0.6507324402580079},{"cp_plan":{"max_rank_tokens":49280,"summary_ms":7.63142853975296,"wave_count":1},"fallback_work":3762914080,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":10671.984375,"n":8,"spread_pct":0.33103613872185794},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.849655717611313,"wave_count":2},"fallback_work":503614624,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":2147.1971435546875,"n":8,"spread_pct":0.7166410107562667},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.7876259088516235,"wave_count":2},"fallback_work":503499136,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":2148.7705078125,"n":8,"spread_pct":0.7178205559956389},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.64722903072834,"wave_count":2},"fallback_work":500018368,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":2149.1783447265625,"n":8,"spread_pct":0.6508664711015275},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.619639530777931,"wave_count":2},"fallback_work":494157280,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":2145.8858642578125,"n":8,"spread_pct":0.23752073272023017},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.387727335095406,"wave_count":2},"fallback_work":490594432,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":2032.5725708007812,"n":8,"spread_pct":0.5008217456007985},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.296753883361816,"wave_count":2},"fallback_work":490684064,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":2029.589111328125,"n":8,"spread_pct":0.4966861726000238}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.5609379410743713,"wave_count":2},"fallback_work":501902752,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1309.0148315429688,"n":8,"spread_pct":0.9407980146686941},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":6.833253428339958,"wave_count":3},"fallback_work":982692416,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":1901.5645751953125,"n":8,"spread_pct":1.1411629904625336},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":10.235555469989777,"wave_count":1},"fallback_work":1468814048,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":2485.478759765625,"n":8,"spread_pct":0.6871455792881327},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.558693453669548,"wave_count":2},"fallback_work":500278048,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1311.9367065429688,"n":8,"spread_pct":1.3895658623206546},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":5.919948220252991,"wave_count":1},"fallback_work":1954919744,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2887.5787353515625,"n":8,"spread_pct":0.6668683149831252},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.514627158641815,"wave_count":2},"fallback_work":498637408,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1309.3713989257812,"n":8,"spread_pct":1.2484092359106933},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.513764753937721,"wave_count":2},"fallback_work":498858368,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1311.3014526367188,"n":8,"spread_pct":1.5001049833313653}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.551192581653595,"wave_count":2},"fallback_work":489795456,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1296.4798583984375,"n":8,"spread_pct":1.9124049750474614},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":8.695261552929878,"wave_count":3},"fallback_work":973686400,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1841.9691772460938,"n":8,"spread_pct":0.8075927944572259}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.3752769231796265,"wave_count":2},"fallback_work":498081696,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1268.148193359375,"n":8,"spread_pct":1.1377009761596046}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":8192,"summary_ms":6.403544917702675,"wave_count":1},"fallback_work":850645248,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":1765.0755004882812,"n":8,"spread_pct":1.4851275789405836},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":3.8229525089263916,"wave_count":2},"fallback_work":664571424,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":1551.6672973632812,"n":8,"spread_pct":1.285089120011209},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.7399772554636,"wave_count":2},"fallback_work":608272224,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1424.3728637695312,"n":8,"spread_pct":56.554871732244074},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.244567826390266,"wave_count":2},"fallback_work":902833312,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":1643.6456298828125,"n":8,"spread_pct":0.8132868465017348},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":4.527771845459938,"wave_count":1},"fallback_work":1129815968,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":2265.961181640625,"n":8,"spread_pct":0.730484280511394},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":4.996869713068008,"wave_count":2},"fallback_work":1447099808,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":2028.7633056640625,"n":8,"spread_pct":0.7267490748439691},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":7.977182045578957,"wave_count":1},"fallback_work":1894158496,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":2954.998291015625,"n":8,"spread_pct":1.1777334037362048},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":5.837077274918556,"wave_count":3},"fallback_work":642274432,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1833.988037109375,"n":8,"spread_pct":1.340579059014126},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":10.206015780568123,"wave_count":2},"fallback_work":597336832,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1666.751220703125,"n":8,"spread_pct":0.8201317325747001}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7680,"summary_ms":6.6765230149030685,"wave_count":2},"fallback_work":690182912,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":1661.8751831054688,"n":8,"spread_pct":0.9765770830738747},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":5.857847630977631,"wave_count":4},"fallback_work":779228704,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":1520.3231201171875,"n":8,"spread_pct":0.7402633283891983},{"cp_plan":{"max_rank_tokens":12288,"summary_ms":9.03952307999134,"wave_count":1},"fallback_work":1240683840,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":2535.751708984375,"n":8,"spread_pct":0.4418838833983428},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.536434844136238,"wave_count":2},"fallback_work":504191232,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":1307.8300170898438,"n":8,"spread_pct":0.8009992025663113}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.036147475242615,"wave_count":4},"fallback_work":537980224,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1654.2457885742188,"n":8,"spread_pct":0.8622019503775261},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":4.247259348630905,"wave_count":2},"fallback_work":1027465056,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":1974.4723510742188,"n":8,"spread_pct":0.6862803134057739},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":8.215250447392464,"wave_count":2},"fallback_work":1507516320,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":2487.7847900390625,"n":8,"spread_pct":0.3839364928236025},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.259564310312271,"wave_count":1},"fallback_work":2006845216,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":3044.132080078125,"n":8,"spread_pct":0.3558251358494606},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.63864004611969,"wave_count":2},"fallback_work":531064288,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1425.6590576171875,"n":8,"spread_pct":0.9960528295275991},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.1149333119392395,"wave_count":2},"fallback_work":523804512,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1315.28857421875,"n":8,"spread_pct":0.9702321444282628},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":16.016777604818344,"wave_count":3},"fallback_work":2505830048,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":4007.4234619140625,"n":8,"spread_pct":0.2089505947108109},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":3.685988485813141,"wave_count":2},"fallback_work":535822144,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1426.217041015625,"n":8,"spread_pct":44.79782292569822},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.218753427267075,"wave_count":2},"fallback_work":523369216,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1319.0337524414062,"n":8,"spread_pct":1.3676244961547215},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.212239757180214,"wave_count":2},"fallback_work":523754016,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1318.3880615234375,"n":8,"spread_pct":0.5823206624670667}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":4096,"summary_ms":3.7152357399463654,"wave_count":1},"fallback_work":374930912,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":914.6889953613281,"n":8,"spread_pct":1.0211750900299406},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6074835807085037,"wave_count":2},"fallback_work":267737312,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":754.4541931152344,"n":8,"spread_pct":1.9803126411321437},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.638271242380142,"wave_count":2},"fallback_work":269377952,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":758.1194763183594,"n":8,"spread_pct":2.172090485736307},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.610735759139061,"wave_count":2},"fallback_work":278180512,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":758.8063354492188,"n":8,"spread_pct":1.548074476747352},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.616565838456154,"wave_count":2},"fallback_work":281748960,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":755.7190551757812,"n":8,"spread_pct":1.2724745977955694},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6275330930948257,"wave_count":2},"fallback_work":283553664,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":756.7072143554688,"n":8,"spread_pct":1.9689558490230314},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.5602282732725143,"wave_count":1},"fallback_work":293012480,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":818.7517700195312,"n":8,"spread_pct":2.144720434988325},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6075711250305176,"wave_count":2},"fallback_work":303307552,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":828.6815795898438,"n":8,"spread_pct":0.5596769093649396},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.629116341471672,"wave_count":2},"fallback_work":306424768,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":829.6932067871094,"n":8,"spread_pct":2.429148487127505},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.5893935710191727,"wave_count":1},"fallback_work":318262624,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":820.0573425292969,"n":8,"spread_pct":1.0757438353499762},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.8856733590364456,"wave_count":2},"fallback_work":253271520,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":758.3445739746094,"n":8,"spread_pct":1.5869576109074512},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.581484779715538,"wave_count":1},"fallback_work":321215776,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":821.2640075683594,"n":8,"spread_pct":0.9436970412927653},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.8391873240470886,"wave_count":2},"fallback_work":254322112,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":757.1726379394531,"n":8,"spread_pct":1.1665858491260723},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.7086810916662216,"wave_count":2},"fallback_work":256669504,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":755.6051635742188,"n":8,"spread_pct":2.2087531861288943},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.58453206717968,"wave_count":2},"fallback_work":332618112,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":830.9009704589844,"n":8,"spread_pct":0.8972061996044142},{"cp_plan":{"max_rank_tokens":4608,"summary_ms":3.8560666143894196,"wave_count":3},"fallback_work":551422816,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":1083.0325927734375,"n":8,"spread_pct":55.149166447160866},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.065794870257378,"wave_count":1},"fallback_work":724961344,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":1257.9899291992188,"n":8,"spread_pct":1.517992391116024},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.367507994174957,"wave_count":1},"fallback_work":908467040,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":1550.2900390625,"n":8,"spread_pct":0.8740567698710612},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":4.847746342420578,"wave_count":3},"fallback_work":1097796896,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":1659.4286499023438,"n":8,"spread_pct":0.7849175368366409},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":5.09238988161087,"wave_count":1},"fallback_work":1289669632,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":1871.4909057617188,"n":8,"spread_pct":0.5807026803124741},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.400395020842552,"wave_count":1},"fallback_work":1479409760,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":2071.0579833984375,"n":8,"spread_pct":0.7595142480240163},{"cp_plan":{"max_rank_tokens":12288,"summary_ms":5.795372650027275,"wave_count":3},"fallback_work":1676122496,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":2453.3663330078125,"n":8,"spread_pct":0.7085986813498691},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":9.894050657749176,"wave_count":3},"fallback_work":1870538336,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":2599.8907470703125,"n":8,"spread_pct":0.25719417578065484},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.7073870450258255,"wave_count":3},"fallback_work":2067579200,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":2776.04150390625,"n":8,"spread_pct":1.025841406299332},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":6.898220628499985,"wave_count":1},"fallback_work":2273069248,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":2984.40576171875,"n":8,"spread_pct":0.516773141899879},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.559831529855728,"wave_count":1},"fallback_work":348819488,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":820.86328125,"n":8,"spread_pct":1.1349819882840568},{"cp_plan":{"max_rank_tokens":16896,"summary_ms":7.33998604118824,"wave_count":1},"fallback_work":2483358336,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":3359.2412109375,"n":8,"spread_pct":0.9338380627368008},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.779226914048195,"wave_count":1},"fallback_work":261206400,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":731.5633850097656,"n":8,"spread_pct":89.84857209089537},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.706619143486023,"wave_count":2},"fallback_work":264102656,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":755.00439453125,"n":8,"spread_pct":3.7447801098012024},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6730878055095673,"wave_count":2},"fallback_work":264841056,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":756.9075622558594,"n":8,"spread_pct":4.3957452897502},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.644118085503578,"wave_count":2},"fallback_work":265292288,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":761.7616271972656,"n":8,"spread_pct":3.5255868009032345},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.669207915663719,"wave_count":2},"fallback_work":264348864,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":760.1323547363281,"n":8,"spread_pct":2.9993796776355177},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":6.277946755290031,"wave_count":1},"fallback_work":2703557472,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":3680.9940185546875,"n":8,"spread_pct":0.6693881216472269},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.920895978808403,"wave_count":1},"fallback_work":263386816,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":732.0968933105469,"n":8,"spread_pct":1.572456045184382},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.997344523668289,"wave_count":2},"fallback_work":257168096,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":760.1166687011719,"n":8,"spread_pct":10.198806359674032},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.961605951189995,"wave_count":2},"fallback_work":252426176,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":761.7869567871094,"n":8,"spread_pct":5.903141285225462},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":3.7991963326931,"wave_count":1},"fallback_work":357761088,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":913.3843994140625,"n":8,"spread_pct":1.4612803702533061},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6427881568670273,"wave_count":2},"fallback_work":336038432,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":831.3416137695312,"n":8,"spread_pct":1.4593889110406286},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.6695022135972977,"wave_count":1},"fallback_work":322067744,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":821.9958801269531,"n":8,"spread_pct":1.7473071722125588},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.697548061609268,"wave_count":2},"fallback_work":312321760,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":828.3391723632812,"n":8,"spread_pct":3.4016482618108275},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.733450546860695,"wave_count":2},"fallback_work":303847328,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":830.5552368164062,"n":8,"spread_pct":2.0542898247382366},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.7912335246801376,"wave_count":2},"fallback_work":291968512,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":831.5709533691406,"n":8,"spread_pct":1.1789312580753208},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":4.31065633893013,"wave_count":2},"fallback_work":285011616,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":756.936279296875,"n":8,"spread_pct":2.5186257563916525},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.8651730865240097,"wave_count":2},"fallback_work":274281248,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":759.5711059570312,"n":8,"spread_pct":3.431590334061267}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7789568305015564,"wave_count":2},"fallback_work":506388704,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1323.1082763671875,"n":8,"spread_pct":46.254118151993794},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":6.0199592262506485,"wave_count":3},"fallback_work":968721056,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":1748.9330444335938,"n":8,"spread_pct":1.187037897931562},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":8.177191019058228,"wave_count":2},"fallback_work":1433801536,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":2311.6439208984375,"n":8,"spread_pct":0.4961929793794546},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":5.954163148999214,"wave_count":1},"fallback_work":1895600736,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":2886.3580322265625,"n":8,"spread_pct":0.2787222439204493},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":11.647026985883713,"wave_count":3},"fallback_work":2355964320,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":3739.9322509765625,"n":8,"spread_pct":0.297693759531816},{"cp_plan":{"max_rank_tokens":19456,"summary_ms":8.304476737976074,"wave_count":1},"fallback_work":2820675488,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":4221.732666015625,"n":8,"spread_pct":0.26208322507171866},{"cp_plan":{"max_rank_tokens":21504,"summary_ms":9.011587128043175,"wave_count":1},"fallback_work":3290965056,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":4702.496337890625,"n":8,"spread_pct":0.39339768475073883},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.765510395169258,"wave_count":2},"fallback_work":491645280,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":1322.9910888671875,"n":8,"spread_pct":46.97405217945877},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.707094117999077,"wave_count":2},"fallback_work":501154592,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1320.542724609375,"n":8,"spread_pct":48.21513186587472},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.755338490009308,"wave_count":2},"fallback_work":491752448,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1319.6349487304688,"n":8,"spread_pct":0.9513202943133863},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.304124042391777,"wave_count":2},"fallback_work":491496928,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":1321.8689575195312,"n":8,"spread_pct":1.0904212817536059},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7101171910762787,"wave_count":2},"fallback_work":492803040,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1321.9671020507812,"n":8,"spread_pct":1.6381467954198856},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.575256258249283,"wave_count":2},"fallback_work":494525824,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1318.0465087890625,"n":8,"spread_pct":0.6305848323078894},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.608025774359703,"wave_count":2},"fallback_work":496232672,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1322.2338256835938,"n":8,"spread_pct":0.9105278058062605},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.75240296125412,"wave_count":2},"fallback_work":491276192,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":1319.60888671875,"n":8,"spread_pct":1.4375264332804576},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.566427320241928,"wave_count":2},"fallback_work":495904544,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1320.7924194335938,"n":8,"spread_pct":1.190608276956282},{"cp_plan":{"max_rank_tokens":24576,"summary_ms":8.50585289299488,"wave_count":1},"fallback_work":3761238688,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5363.1201171875,"n":8,"spread_pct":0.14565259152906088},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7470124661922455,"wave_count":2},"fallback_work":504107744,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1319.9102783203125,"n":8,"spread_pct":0.6362700175395518},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7496164441108704,"wave_count":2},"fallback_work":504123456,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1322.0388793945312,"n":8,"spread_pct":1.7360713397815402},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.71385195851326,"wave_count":2},"fallback_work":500775968,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1322.50927734375,"n":8,"spread_pct":1.217953191047883},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.714174196124077,"wave_count":2},"fallback_work":495049664,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":1322.0695190429688,"n":8,"spread_pct":0.7274712594699352},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.850731998682022,"wave_count":2},"fallback_work":491620032,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":1320.1898193359375,"n":8,"spread_pct":1.0311424092197916},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.872191533446312,"wave_count":2},"fallback_work":491840992,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":1323.8377075195312,"n":8,"spread_pct":1.5313311255630941}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12204,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":502191616,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1520.9679565429688,"n":8,"spread_pct":0.26116855532520084},{"cp_plan":{"max_rank_tokens":23926,"summary_ms":0.00018067657947540283,"wave_count":0},"fallback_work":983825152,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":2893.6900634765625,"n":8,"spread_pct":0.5095104181971912},{"cp_plan":{"max_rank_tokens":35778,"summary_ms":0.00021047890186309814,"wave_count":0},"fallback_work":1470800128,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":4238.941162109375,"n":8,"spread_pct":0.4306702706370086},{"cp_plan":{"max_rank_tokens":12154,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":500366592,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1506.4594116210938,"n":8,"spread_pct":1.027217146086123},{"cp_plan":{"max_rank_tokens":47640,"summary_ms":0.000303611159324646,"wave_count":0},"fallback_work":1957956608,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5755.831787109375,"n":8,"spread_pct":0.18690282999744612},{"cp_plan":{"max_rank_tokens":12114,"summary_ms":0.00021420419216156006,"wave_count":0},"fallback_work":498723072,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1505.9714965820312,"n":8,"spread_pct":36.30366043189797},{"cp_plan":{"max_rank_tokens":12109,"summary_ms":0.00021047890186309814,"wave_count":0},"fallback_work":498747008,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1504.1402587890625,"n":8,"spread_pct":1.0481080176079745}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11920,"summary_ms":0.0002551823854446411,"wave_count":0},"fallback_work":490260480,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1471.9541625976562,"n":8,"spread_pct":0.49376159398274133},{"cp_plan":{"max_rank_tokens":23728,"summary_ms":0.00026263296604156494,"wave_count":0},"fallback_work":975198208,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2861.1475830078125,"n":8,"spread_pct":0.46273397540740735}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12138,"summary_ms":0.00024959444999694824,"wave_count":0},"fallback_work":498857216,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1504.901611328125,"n":8,"spread_pct":0.693112691486495}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":20701,"summary_ms":0.00031851232051849365,"wave_count":0},"fallback_work":851447424,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":2561.71337890625,"n":8,"spread_pct":0.6466257077010492},{"cp_plan":{"max_rank_tokens":16154,"summary_ms":0.00028870999813079834,"wave_count":0},"fallback_work":664849664,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":1994.7476196289062,"n":8,"spread_pct":0.2678604348670191},{"cp_plan":{"max_rank_tokens":14771,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":608254336,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1826.611328125,"n":8,"spread_pct":32.08099655163196},{"cp_plan":{"max_rank_tokens":21963,"summary_ms":0.00017881393432617188,"wave_count":0},"fallback_work":903529856,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":2624.9617919921875,"n":8,"spread_pct":0.6093094640002898},{"cp_plan":{"max_rank_tokens":27497,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":1130910848,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":3217.82958984375,"n":8,"spread_pct":0.4069276484553001},{"cp_plan":{"max_rank_tokens":35243,"summary_ms":0.0001862645149230957,"wave_count":0},"fallback_work":1448949120,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":4087.6319580078125,"n":8,"spread_pct":0.559208819979979},{"cp_plan":{"max_rank_tokens":46153,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":1896989824,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":5412.927734375,"n":8,"spread_pct":0.6209093608725349},{"cp_plan":{"max_rank_tokens":15600,"summary_ms":0.0002868473529815674,"wave_count":0},"fallback_work":642316288,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1924.72705078125,"n":8,"spread_pct":0.8019666445196546},{"cp_plan":{"max_rank_tokens":14494,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":597102336,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1796.3043212890625,"n":8,"spread_pct":0.7565582139708094}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16800,"summary_ms":0.00032223761081695557,"wave_count":0},"fallback_work":690900992,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":2082.9786376953125,"n":8,"spread_pct":0.5733437077558693},{"cp_plan":{"max_rank_tokens":18971,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":780103040,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":2249.3695068359375,"n":8,"spread_pct":0.676643601573466},{"cp_plan":{"max_rank_tokens":30232,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":1242565632,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":3595.2459716796875,"n":8,"spread_pct":0.9320440828724319},{"cp_plan":{"max_rank_tokens":12255,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":504385408,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":1515.5989990234375,"n":8,"spread_pct":0.5799310184686641}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13078,"summary_ms":0.0002514570951461792,"wave_count":0},"fallback_work":538233600,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1659.2310180664062,"n":8,"spread_pct":26.849868277691968},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.00023283064365386963,"wave_count":0},"fallback_work":1028577792,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":3052.7406005859375,"n":8,"spread_pct":0.5113153360091915},{"cp_plan":{"max_rank_tokens":36716,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":1509471744,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":4426.281982421875,"n":8,"spread_pct":0.566606238449081},{"cp_plan":{"max_rank_tokens":48890,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":2009677056,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":5908.1591796875,"n":8,"spread_pct":0.45197856843563294},{"cp_plan":{"max_rank_tokens":12899,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":531108224,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1633.7697143554688,"n":8,"spread_pct":0.9127797874773099},{"cp_plan":{"max_rank_tokens":12722,"summary_ms":0.00027008354663848877,"wave_count":0},"fallback_work":523835648,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1607.5345458984375,"n":8,"spread_pct":0.26321059056993495},{"cp_plan":{"max_rank_tokens":61066,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":2509735168,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":7427.224853515625,"n":8,"spread_pct":0.24256857219991393},{"cp_plan":{"max_rank_tokens":13015,"summary_ms":0.00021420419216156006,"wave_count":0},"fallback_work":535874432,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1662.8590698242188,"n":8,"spread_pct":0.7620314324293284},{"cp_plan":{"max_rank_tokens":12701,"summary_ms":0.0002384185791015625,"wave_count":0},"fallback_work":523202176,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1605.537353515625,"n":8,"spread_pct":0.3750600073202468},{"cp_plan":{"max_rank_tokens":12700,"summary_ms":0.00027567148208618164,"wave_count":0},"fallback_work":523390464,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1603.0416870117188,"n":8,"spread_pct":0.3979556229221982}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9058,"summary_ms":0.00030919909477233887,"wave_count":0},"fallback_work":374108416,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1099.8445434570312,"n":8,"spread_pct":0.9447246878207053},{"cp_plan":{"max_rank_tokens":6403,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":265937280,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":810.928955078125,"n":8,"spread_pct":0.32809835122901404},{"cp_plan":{"max_rank_tokens":6443,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":267580800,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":812.9432678222656,"n":8,"spread_pct":0.7616337616290692},{"cp_plan":{"max_rank_tokens":6668,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":276596224,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":832.3840942382812,"n":8,"spread_pct":1.0006319943105115},{"cp_plan":{"max_rank_tokens":6755,"summary_ms":0.00019744038581848145,"wave_count":0},"fallback_work":280170880,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":845.4008178710938,"n":8,"spread_pct":0.4269207952294148},{"cp_plan":{"max_rank_tokens":6799,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":281978752,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":854.7799987792969,"n":8,"spread_pct":0.4737618554264812},{"cp_plan":{"max_rank_tokens":7040,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":291651584,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":880.9264526367188,"n":8,"spread_pct":0.9865033143609699},{"cp_plan":{"max_rank_tokens":7291,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":301964672,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":893.355712890625,"n":8,"spread_pct":2.37615070022423},{"cp_plan":{"max_rank_tokens":7367,"summary_ms":0.00017508864402770996,"wave_count":0},"fallback_work":305087360,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":910.5963134765625,"n":8,"spread_pct":0.6711012178258768},{"cp_plan":{"max_rank_tokens":7666,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":317143296,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":943.1966552734375,"n":8,"spread_pct":0.8198166169190415},{"cp_plan":{"max_rank_tokens":5988,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":250262016,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":758.3076171875,"n":8,"spread_pct":0.5089941236846545},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.00018440186977386475,"wave_count":0},"fallback_work":320101632,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":954.04931640625,"n":8,"spread_pct":0.7291276953241554},{"cp_plan":{"max_rank_tokens":6024,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":251511808,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":759.2486267089844,"n":8,"spread_pct":3.0334587815399288},{"cp_plan":{"max_rank_tokens":6102,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":254257920,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":775.2062683105469,"n":8,"spread_pct":1.0919472785543825},{"cp_plan":{"max_rank_tokens":8016,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":331524096,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":993.094482421875,"n":8,"spread_pct":3.1230555729690996},{"cp_plan":{"max_rank_tokens":13361,"summary_ms":0.00017695128917694092,"wave_count":0},"fallback_work":550910080,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":1570.3699340820312,"n":8,"spread_pct":0.4511417706620727},{"cp_plan":{"max_rank_tokens":17592,"summary_ms":0.00019371509552001953,"wave_count":0},"fallback_work":724753408,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":1989.6441650390625,"n":8,"spread_pct":0.4514464391274032},{"cp_plan":{"max_rank_tokens":22066,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":908581120,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":2494.1470947265625,"n":8,"spread_pct":0.8833277791517083},{"cp_plan":{"max_rank_tokens":26682,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":1098243328,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":2974.7947998046875,"n":8,"spread_pct":0.8078042081995955},{"cp_plan":{"max_rank_tokens":31360,"summary_ms":0.0002514570951461792,"wave_count":0},"fallback_work":1290452992,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":3530.005615234375,"n":8,"spread_pct":0.7973564799487208},{"cp_plan":{"max_rank_tokens":35986,"summary_ms":0.00017695128917694092,"wave_count":0},"fallback_work":1480526080,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":3970.5367431640625,"n":8,"spread_pct":0.4787891228964125},{"cp_plan":{"max_rank_tokens":40782,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":1677584128,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":4465.0498046875,"n":8,"spread_pct":0.728017582740604},{"cp_plan":{"max_rank_tokens":45522,"summary_ms":0.00017695128917694092,"wave_count":0},"fallback_work":1872341248,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":5003.5234375,"n":8,"spread_pct":0.6297320169692919},{"cp_plan":{"max_rank_tokens":50326,"summary_ms":0.00018067657947540283,"wave_count":0},"fallback_work":2069728000,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":5492.16455078125,"n":8,"spread_pct":6.04253947362689},{"cp_plan":{"max_rank_tokens":55336,"summary_ms":0.0001825392246246338,"wave_count":0},"fallback_work":2275578880,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":6007.5107421875,"n":8,"spread_pct":0.6651748275392896},{"cp_plan":{"max_rank_tokens":8411,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":347753856,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":1034.8089599609375,"n":8,"spread_pct":0.9965135708855152},{"cp_plan":{"max_rank_tokens":60463,"summary_ms":0.00017136335372924805,"wave_count":0},"fallback_work":2486237056,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":6563.6162109375,"n":8,"spread_pct":0.2838802864334236},{"cp_plan":{"max_rank_tokens":6223,"summary_ms":0.0002905726432800293,"wave_count":0},"fallback_work":259000192,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":784.5099182128906,"n":8,"spread_pct":1.0947911451205887},{"cp_plan":{"max_rank_tokens":6304,"summary_ms":0.00017881393432617188,"wave_count":0},"fallback_work":262098944,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":810.5380249023438,"n":8,"spread_pct":2.1431107493267056},{"cp_plan":{"max_rank_tokens":6322,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":262838528,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":805.4425964355469,"n":8,"spread_pct":0.4985764458055762},{"cp_plan":{"max_rank_tokens":6333,"summary_ms":0.00017136335372924805,"wave_count":0},"fallback_work":263290496,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":805.9715881347656,"n":8,"spread_pct":0.6283586375003753},{"cp_plan":{"max_rank_tokens":6310,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":262345472,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":805.3670959472656,"n":8,"spread_pct":1.0020057014197166},{"cp_plan":{"max_rank_tokens":65842,"summary_ms":0.00037066638469696045,"wave_count":0},"fallback_work":2707020032,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":7275.091064453125,"n":8,"spread_pct":0.23462039863184972},{"cp_plan":{"max_rank_tokens":6245,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":260592256,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":788.0769958496094,"n":8,"spread_pct":3.341294106225725},{"cp_plan":{"max_rank_tokens":6083,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":254165376,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":774.1786193847656,"n":8,"spread_pct":0.6037459765214716},{"cp_plan":{"max_rank_tokens":5957,"summary_ms":0.00019371509552001953,"wave_count":0},"fallback_work":249217664,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":757.4954223632812,"n":8,"spread_pct":0.6146096083211363},{"cp_plan":{"max_rank_tokens":8629,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":356711040,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1060.4503173828125,"n":8,"spread_pct":1.0759957542174938},{"cp_plan":{"max_rank_tokens":8089,"summary_ms":0.00023096799850463867,"wave_count":0},"fallback_work":334752896,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1002.2061157226562,"n":8,"spread_pct":0.9920557948787906},{"cp_plan":{"max_rank_tokens":7738,"summary_ms":0.00023283064365386963,"wave_count":0},"fallback_work":320560384,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":957.3031311035156,"n":8,"spread_pct":0.9350086099583509},{"cp_plan":{"max_rank_tokens":7490,"summary_ms":0.0003073364496231079,"wave_count":0},"fallback_work":310599936,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":928.2797241210938,"n":8,"spread_pct":0.9033900138306856},{"cp_plan":{"max_rank_tokens":7273,"summary_ms":0.00023469328880310059,"wave_count":0},"fallback_work":301913216,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":895.5096740722656,"n":8,"spread_pct":3.1285853720110524},{"cp_plan":{"max_rank_tokens":6973,"summary_ms":0.00020116567611694336,"wave_count":0},"fallback_work":289816192,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":863.35546875,"n":8,"spread_pct":1.5585764119826802},{"cp_plan":{"max_rank_tokens":6793,"summary_ms":0.00022351741790771484,"wave_count":0},"fallback_work":282649728,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":848.7398071289062,"n":8,"spread_pct":1.00229739446688},{"cp_plan":{"max_rank_tokens":6521,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":271703168,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":821.7198486328125,"n":8,"spread_pct":0.9887426764409691}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12291,"summary_ms":0.0003296881914138794,"wave_count":0},"fallback_work":506290560,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1571.973876953125,"n":8,"spread_pct":0.18977900402326206},{"cp_plan":{"max_rank_tokens":23563,"summary_ms":0.00018440186977386475,"wave_count":0},"fallback_work":969434496,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":2879.6141357421875,"n":8,"spread_pct":0.8103078792659488},{"cp_plan":{"max_rank_tokens":34902,"summary_ms":0.0001862645149230957,"wave_count":0},"fallback_work":1435331328,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":4218.843505859375,"n":8,"spread_pct":0.442953145193599},{"cp_plan":{"max_rank_tokens":46161,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":1897941120,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":5502.14208984375,"n":8,"spread_pct":0.2271221249323439},{"cp_plan":{"max_rank_tokens":57385,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":2359112832,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":6847.028564453125,"n":8,"spread_pct":0.4408276300569337},{"cp_plan":{"max_rank_tokens":68715,"summary_ms":0.00019185245037078857,"wave_count":0},"fallback_work":2824639872,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":8252.5986328125,"n":8,"spread_pct":0.18200943097823946},{"cp_plan":{"max_rank_tokens":80181,"summary_ms":0.0003557652235031128,"wave_count":0},"fallback_work":3295754880,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":9576.09423828125,"n":8,"spread_pct":0.4097218613999573},{"cp_plan":{"max_rank_tokens":11890,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":490731776,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":1534.2402954101562,"n":8,"spread_pct":0.8620840209487269},{"cp_plan":{"max_rank_tokens":12153,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":500849792,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1549.9888916015625,"n":8,"spread_pct":0.6576014240273427},{"cp_plan":{"max_rank_tokens":11903,"summary_ms":0.00022910535335540771,"wave_count":0},"fallback_work":491036544,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1542.1856079101562,"n":8,"spread_pct":0.2977064822734989},{"cp_plan":{"max_rank_tokens":11876,"summary_ms":0.00023096799850463867,"wave_count":0},"fallback_work":490385920,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":1531.768310546875,"n":8,"spread_pct":0.6711695005023013},{"cp_plan":{"max_rank_tokens":11939,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":492286336,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1532.8385009765625,"n":8,"spread_pct":0.4326904814066848},{"cp_plan":{"max_rank_tokens":11981,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":494012032,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1541.0689697265625,"n":8,"spread_pct":1.0622414290633315},{"cp_plan":{"max_rank_tokens":12033,"summary_ms":0.00021047890186309814,"wave_count":0},"fallback_work":495919232,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1540.790771484375,"n":8,"spread_pct":0.8147259763346331},{"cp_plan":{"max_rank_tokens":11881,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":490361984,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":1530.9385986328125,"n":8,"spread_pct":1.5927505665413941},{"cp_plan":{"max_rank_tokens":12025,"summary_ms":0.00023469328880310059,"wave_count":0},"fallback_work":495590528,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1542.2467651367188,"n":8,"spread_pct":0.38957311156395136},{"cp_plan":{"max_rank_tokens":91657,"summary_ms":0.00043958425521850586,"wave_count":0},"fallback_work":3767051392,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":11057.974609375,"n":8,"spread_pct":0.3859982976793296},{"cp_plan":{"max_rank_tokens":12225,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":503808128,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1566.9703979492188,"n":8,"spread_pct":0.934186616003284},{"cp_plan":{"max_rank_tokens":12215,"summary_ms":0.0002216547727584839,"wave_count":0},"fallback_work":503626624,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1565.2771606445312,"n":8,"spread_pct":0.459043479010217},{"cp_plan":{"max_rank_tokens":12123,"summary_ms":0.00024028122425079346,"wave_count":0},"fallback_work":500075904,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1545.7000732421875,"n":8,"spread_pct":0.8402373172489362},{"cp_plan":{"max_rank_tokens":11973,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":494142080,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":1538.1080322265625,"n":8,"spread_pct":0.654855672311374},{"cp_plan":{"max_rank_tokens":11879,"summary_ms":0.00024028122425079346,"wave_count":0},"fallback_work":490509184,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":1528.33837890625,"n":8,"spread_pct":0.7009901160912794},{"cp_plan":{"max_rank_tokens":11874,"summary_ms":0.00021420419216156006,"wave_count":0},"fallback_work":490533120,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":1534.5230712890625,"n":8,"spread_pct":1.071752479334471}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.678605288267136,"wave_count":2},"fallback_work":501802432,"features":{"max_depth":2,"packed_tokens":12204,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1294.2112426757812,"n":8,"spread_pct":0.5603942623678709},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":12.101059779524803,"wave_count":1},"fallback_work":982873312,"features":{"max_depth":2,"packed_tokens":23926,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_11813","median_ms":1958.0827026367188,"n":8,"spread_pct":0.7273848970926206},{"cp_plan":{"max_rank_tokens":20480,"summary_ms":6.127811968326569,"wave_count":1},"fallback_work":1469279392,"features":{"max_depth":2,"packed_tokens":35778,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11858","median_ms":2606.904296875,"n":8,"spread_pct":0.4263674062641648},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.377732217311859,"wave_count":2},"fallback_work":500045344,"features":{"max_depth":3,"packed_tokens":12154,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1292.6282958984375,"n":8,"spread_pct":0.5953701366419085},{"cp_plan":{"max_rank_tokens":27136,"summary_ms":9.40418615937233,"wave_count":1},"fallback_work":1955800960,"features":{"max_depth":1,"packed_tokens":47640,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3449.916259765625,"n":8,"spread_pct":0.6715799331915963},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.360174924135208,"wave_count":2},"fallback_work":498403744,"features":{"max_depth":3,"packed_tokens":12114,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1296.4439697265625,"n":8,"spread_pct":0.7680828407860332},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.413511767983437,"wave_count":2},"fallback_work":498493536,"features":{"max_depth":4,"packed_tokens":12109,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1294.2615356445312,"n":8,"spread_pct":0.7588051240873633}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"9d7ce202a0f10b20cc44497c97c8a58cfdf7d42405c6f1143471483ef41ef0ac","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":0,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.186529994010925,"wave_count":2},"fallback_work":489819392,"features":{"max_depth":2,"packed_tokens":11920,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1280.1307983398438,"n":8,"spread_pct":1.1651279035045048},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":8.319804444909096,"wave_count":1},"fallback_work":974124800,"features":{"max_depth":1,"packed_tokens":23728,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1945.5358276367188,"n":8,"spread_pct":0.7142817031641604}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"f76f71ae0335714b0b4d31c31ae0b34cc53b5a64327eee3aee39f95b1cea8b22","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":1,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.1458608359098434,"wave_count":2},"fallback_work":498307360,"features":{"max_depth":1,"packed_tokens":12138,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1237.9102172851562,"n":8,"spread_pct":0.8515662791507768}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"9a85786ed613f8815b2288ad1a865843ed0747baa545ff87e14c3f7810f77856","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":2,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13824,"summary_ms":8.453572168946266,"wave_count":1},"fallback_work":850683232,"features":{"max_depth":2,"packed_tokens":20701,"segment_count":6,"segments_below":[0,0,0,0,1,2,3,6]},"label":"depth_one","median_ms":1823.715576171875,"n":8,"spread_pct":0.6477696805823984},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.264576688408852,"wave_count":2},"fallback_work":664369184,"features":{"max_depth":3,"packed_tokens":16154,"segment_count":7,"segments_below":[0,0,1,2,3,4,5,7]},"label":"minimum_effective_span_1384","median_ms":1510.8218994140625,"n":8,"spread_pct":1.0650456145130673},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":5.875308066606522,"wave_count":2},"fallback_work":607905856,"features":{"max_depth":3,"packed_tokens":14771,"segment_count":8,"segments_below":[0,0,2,3,5,6,6,8]},"label":"minimum_effective_span_278","median_ms":1426.46484375,"n":8,"spread_pct":1.1576127883891285},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.070785224437714,"wave_count":1},"fallback_work":902770624,"features":{"max_depth":2,"packed_tokens":21963,"segment_count":7,"segments_below":[0,0,1,2,3,5,5,6]},"label":"minimum_effective_span_6364","median_ms":1485.4303588867188,"n":8,"spread_pct":0.7777776088554802},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":7.482681423425674,"wave_count":4},"fallback_work":1129885984,"features":{"max_depth":2,"packed_tokens":27497,"segment_count":7,"segments_below":[0,0,2,3,4,4,4,6]},"label":"minimum_effective_span_6641","median_ms":1801.3089599609375,"n":8,"spread_pct":0.7147986663073179},{"cp_plan":{"max_rank_tokens":17835,"summary_ms":6.4641982316970825,"wave_count":1},"fallback_work":1447486912,"features":{"max_depth":2,"packed_tokens":35243,"segment_count":6,"segments_below":[0,0,1,2,2,2,2,4]},"label":"minimum_effective_span_7747","median_ms":2229.1429443359375,"n":8,"spread_pct":0.7918892682433445},{"cp_plan":{"max_rank_tokens":24649,"summary_ms":7.178476080298424,"wave_count":1},"fallback_work":1894938400,"features":{"max_depth":1,"packed_tokens":46153,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,2]},"label":"no_sharing","median_ms":3074.0965576171875,"n":8,"spread_pct":0.6286781014445499},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":6.883576512336731,"wave_count":3},"fallback_work":641927936,"features":{"max_depth":3,"packed_tokens":15600,"segment_count":8,"segments_below":[0,0,1,3,4,6,6,8]},"label":"uniform_depth_2","median_ms":1427.9166259765625,"n":8,"spread_pct":1.274298689471325},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":7.2480011731386185,"wave_count":3},"fallback_work":596832608,"features":{"max_depth":4,"packed_tokens":14494,"segment_count":9,"segments_below":[0,0,2,4,6,7,7,9]},"label":"uniform_depth_3","median_ms":1371.1220703125,"n":8,"spread_pct":0.9014058826420251}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"73eae660c41accc801ed6f77e31d387eaecd03f3e5e053d1e44533c8d8eefe9f","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":3,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11776,"summary_ms":5.885723978281021,"wave_count":2},"fallback_work":690258432,"features":{"max_depth":2,"packed_tokens":16800,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,4]},"label":"depth_one","median_ms":1550.2525634765625,"n":8,"spread_pct":0.6368197470343518},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":6.0383472591638565,"wave_count":1},"fallback_work":779356352,"features":{"max_depth":2,"packed_tokens":18971,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_6717","median_ms":1747.1897583007812,"n":8,"spread_pct":0.5777204234604278},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":8.435318246483803,"wave_count":1},"fallback_work":1241212800,"features":{"max_depth":1,"packed_tokens":30232,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":2242.3265380859375,"n":8,"spread_pct":0.5041390344969353},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.664762109518051,"wave_count":2},"fallback_work":504026624,"features":{"max_depth":3,"packed_tokens":12255,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,5]},"label":"uniform_depth_2","median_ms":1285.4769897460938,"n":8,"spread_pct":0.5313073698599633}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"4c99985df49b9c1d9d0303783f37a479917a9cd2967b242bf10552a33e56b729","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":4,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.584198981523514,"wave_count":2},"fallback_work":537835232,"features":{"max_depth":2,"packed_tokens":13078,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1375.8677978515625,"n":8,"spread_pct":0.9968140668322878},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":8.747696876525879,"wave_count":1},"fallback_work":1027606592,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":6,"segments_below":[0,1,3,4,4,4,4,4]},"label":"minimum_effective_span_11998","median_ms":1732.6022338867188,"n":8,"spread_pct":0.7767511520696743},{"cp_plan":{"max_rank_tokens":21504,"summary_ms":6.983067840337753,"wave_count":1},"fallback_work":1507938752,"features":{"max_depth":2,"packed_tokens":36716,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12019","median_ms":2746.1607666015625,"n":8,"spread_pct":0.9962150194680971},{"cp_plan":{"max_rank_tokens":25850,"summary_ms":8.884694427251816,"wave_count":1},"fallback_work":2007559712,"features":{"max_depth":2,"packed_tokens":48890,"segment_count":6,"segments_below":[1,1,2,2,2,2,2,2]},"label":"minimum_effective_span_12176","median_ms":3383.3212890625,"n":8,"spread_pct":0.3234646290208661},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.577540025115013,"wave_count":2},"fallback_work":530784064,"features":{"max_depth":3,"packed_tokens":12899,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"minimum_effective_span_179","median_ms":1371.197509765625,"n":8,"spread_pct":1.725991177661947},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.403734743595123,"wave_count":2},"fallback_work":523519904,"features":{"max_depth":3,"packed_tokens":12722,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_22","median_ms":1373.3851318359375,"n":8,"spread_pct":0.886668122566889},{"cp_plan":{"max_rank_tokens":33280,"summary_ms":7.013779133558273,"wave_count":1},"fallback_work":2506967840,"features":{"max_depth":1,"packed_tokens":61066,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":4203.771728515625,"n":8,"spread_pct":0.47675052787242483},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.64892403781414,"wave_count":2},"fallback_work":535544704,"features":{"max_depth":3,"packed_tokens":13015,"segment_count":7,"segments_below":[1,2,5,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1372.9481811523438,"n":8,"spread_pct":0.6563238476023103},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.47630712389946,"wave_count":2},"fallback_work":522953056,"features":{"max_depth":4,"packed_tokens":12701,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1371.0462036132812,"n":8,"spread_pct":0.7236540120403926},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.4678767919540405,"wave_count":2},"fallback_work":523206848,"features":{"max_depth":5,"packed_tokens":12700,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1372.0048217773438,"n":8,"spread_pct":1.1303564445046688}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"806be478da7bc33eaffa693da4147c77cfc56a4c08d9d4338dd6afee698256c0","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":5,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7680,"summary_ms":6.6099632531404495,"wave_count":1},"fallback_work":374165152,"features":{"max_depth":2,"packed_tokens":9058,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1073.6336669921875,"n":8,"spread_pct":1.9138946928707874},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.045197904109955,"wave_count":2},"fallback_work":266383680,"features":{"max_depth":6,"packed_tokens":6403,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_105","median_ms":931.6465759277344,"n":8,"spread_pct":67.9606049424147},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.0400495529174805,"wave_count":2},"fallback_work":268025280,"features":{"max_depth":6,"packed_tokens":6443,"segment_count":18,"segments_below":[5,11,17,17,17,17,17,18]},"label":"minimum_effective_span_113","median_ms":928.695068359375,"n":8,"spread_pct":3.7165527276359875},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.0811638832092285,"wave_count":2},"fallback_work":276964288,"features":{"max_depth":5,"packed_tokens":6668,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":924.6837768554688,"n":8,"spread_pct":1.338955416491557},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.115281954407692,"wave_count":2},"fallback_work":280534848,"features":{"max_depth":5,"packed_tokens":6755,"segment_count":17,"segments_below":[3,9,16,16,16,16,16,17]},"label":"minimum_effective_span_164","median_ms":926.477294921875,"n":8,"spread_pct":2.797072928555034},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.05740025639534,"wave_count":2},"fallback_work":282340608,"features":{"max_depth":5,"packed_tokens":6799,"segment_count":17,"segments_below":[4,8,16,16,16,16,16,17]},"label":"minimum_effective_span_195","median_ms":924.8861389160156,"n":8,"spread_pct":1.0052968869657037},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.491092965006828,"wave_count":2},"fallback_work":291936256,"features":{"max_depth":4,"packed_tokens":7040,"segment_count":16,"segments_below":[3,6,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":974.6739501953125,"n":8,"spread_pct":0.900823329716756},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.831997841596603,"wave_count":2},"fallback_work":302237376,"features":{"max_depth":4,"packed_tokens":7291,"segment_count":16,"segments_below":[3,7,11,15,15,15,15,16]},"label":"minimum_effective_span_245","median_ms":971.885986328125,"n":8,"spread_pct":1.409254819014556},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.92535175383091,"wave_count":2},"fallback_work":305356416,"features":{"max_depth":4,"packed_tokens":7367,"segment_count":16,"segments_below":[3,6,10,15,15,15,15,16]},"label":"minimum_effective_span_281","median_ms":970.1366577148438,"n":8,"spread_pct":1.251610676476885},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.152046844363213,"wave_count":2},"fallback_work":317332384,"features":{"max_depth":3,"packed_tokens":7666,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_300","median_ms":1028.2507934570312,"n":8,"spread_pct":2.283291825106173},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.566975101828575,"wave_count":1},"fallback_work":251121472,"features":{"max_depth":12,"packed_tokens":5988,"segment_count":24,"segments_below":[16,23,23,23,23,23,23,24]},"label":"minimum_effective_span_32","median_ms":861.1315307617188,"n":8,"spread_pct":3.2470067041237236},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":7.416959851980209,"wave_count":1},"fallback_work":320287264,"features":{"max_depth":3,"packed_tokens":7738,"segment_count":15,"segments_below":[1,4,9,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":1010.0133056640625,"n":8,"spread_pct":0.9352033936550137},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.550622940063477,"wave_count":1},"fallback_work":252304000,"features":{"max_depth":11,"packed_tokens":6024,"segment_count":23,"segments_below":[15,22,22,22,22,22,22,23]},"label":"minimum_effective_span_37","median_ms":862.5213928222656,"n":8,"spread_pct":3.3890407142514194},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.113944411277771,"wave_count":1},"fallback_work":254915296,"features":{"max_depth":9,"packed_tokens":6102,"segment_count":21,"segments_below":[9,20,20,20,20,20,20,21]},"label":"minimum_effective_span_40","median_ms":857.1520080566406,"n":8,"spread_pct":1.5417366734554159},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.854832172393799,"wave_count":1},"fallback_work":331696384,"features":{"max_depth":3,"packed_tokens":8016,"segment_count":15,"segments_below":[1,4,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":1010.0321044921875,"n":8,"spread_pct":57.35235128350437},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":5.650587379932404,"wave_count":2},"fallback_work":550760352,"features":{"max_depth":2,"packed_tokens":13361,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4733","median_ms":1368.7799682617188,"n":8,"spread_pct":1.0473782524114434},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":6.242960691452026,"wave_count":1},"fallback_work":724400512,"features":{"max_depth":2,"packed_tokens":17592,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4772","median_ms":1564.0809326171875,"n":8,"spread_pct":1.2458878520646806},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.944512948393822,"wave_count":1},"fallback_work":908013472,"features":{"max_depth":2,"packed_tokens":22066,"segment_count":14,"segments_below":[2,3,5,8,10,10,10,14]},"label":"minimum_effective_span_4826","median_ms":1652.1439819335938,"n":8,"spread_pct":1.0128366952756958},{"cp_plan":{"max_rank_tokens":16896,"summary_ms":7.4920207262039185,"wave_count":1},"fallback_work":1097454112,"features":{"max_depth":2,"packed_tokens":26682,"segment_count":14,"segments_below":[2,3,5,7,9,9,9,14]},"label":"minimum_effective_span_4865","median_ms":1987.3873901367188,"n":8,"spread_pct":0.9730921824352825},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":8.083287626504898,"wave_count":1},"fallback_work":1289439232,"features":{"max_depth":2,"packed_tokens":31360,"segment_count":14,"segments_below":[1,3,4,7,8,8,8,14]},"label":"minimum_effective_span_4896","median_ms":2178.92431640625,"n":8,"spread_pct":0.6044791220581938},{"cp_plan":{"max_rank_tokens":20992,"summary_ms":8.613333106040955,"wave_count":1},"fallback_work":1479290272,"features":{"max_depth":2,"packed_tokens":35986,"segment_count":14,"segments_below":[2,2,3,7,7,7,7,14]},"label":"minimum_effective_span_4927","median_ms":2470.62548828125,"n":8,"spread_pct":24.158130660627478},{"cp_plan":{"max_rank_tokens":23040,"summary_ms":7.425013929605484,"wave_count":1},"fallback_work":1676118112,"features":{"max_depth":2,"packed_tokens":40782,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_4977","median_ms":2701.658447265625,"n":8,"spread_pct":0.7133205307467663},{"cp_plan":{"max_rank_tokens":26112,"summary_ms":6.935250014066696,"wave_count":2},"fallback_work":1870647712,"features":{"max_depth":2,"packed_tokens":45522,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5013","median_ms":3049.7847900390625,"n":8,"spread_pct":0.7392619476343846},{"cp_plan":{"max_rank_tokens":27648,"summary_ms":8.406182751059532,"wave_count":1},"fallback_work":2067803872,"features":{"max_depth":2,"packed_tokens":50326,"segment_count":14,"segments_below":[0,1,3,4,4,4,4,14]},"label":"minimum_effective_span_5081","median_ms":3242.633056640625,"n":8,"spread_pct":0.565208470974433},{"cp_plan":{"max_rank_tokens":30720,"summary_ms":11.204015463590622,"wave_count":1},"fallback_work":2273414272,"features":{"max_depth":2,"packed_tokens":55336,"segment_count":14,"segments_below":[0,1,3,3,3,3,3,14]},"label":"minimum_effective_span_5173","median_ms":3594.8162841796875,"n":8,"spread_pct":0.9960020354015207},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.296403706073761,"wave_count":1},"fallback_work":347907264,"features":{"max_depth":3,"packed_tokens":8411,"segment_count":15,"segments_below":[2,5,8,11,14,14,14,15]},"label":"minimum_effective_span_522","median_ms":1006.7870178222656,"n":8,"spread_pct":1.1967906903333112},{"cp_plan":{"max_rank_tokens":32768,"summary_ms":9.316431358456612,"wave_count":1},"fallback_work":2483826432,"features":{"max_depth":2,"packed_tokens":60463,"segment_count":14,"segments_below":[1,2,2,2,2,2,2,14]},"label":"minimum_effective_span_5254","median_ms":3809.3450927734375,"n":8,"spread_pct":0.49907462165716987},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.0763397961854935,"wave_count":1},"fallback_work":259586304,"features":{"max_depth":8,"packed_tokens":6223,"segment_count":20,"segments_below":[6,18,19,19,19,19,19,20]},"label":"minimum_effective_span_71","median_ms":914.3616943359375,"n":8,"spread_pct":1.3475223320998964},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.415033549070358,"wave_count":1},"fallback_work":262615552,"features":{"max_depth":7,"packed_tokens":6304,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_82","median_ms":912.2492065429688,"n":8,"spread_pct":1.4495569502828494},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.057560279965401,"wave_count":1},"fallback_work":263354272,"features":{"max_depth":7,"packed_tokens":6322,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_87","median_ms":912.5244140625,"n":8,"spread_pct":3.802070792198411},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.073832347989082,"wave_count":1},"fallback_work":263805792,"features":{"max_depth":7,"packed_tokens":6333,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_93","median_ms":914.0802307128906,"n":8,"spread_pct":1.7402640527038389},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.359956994652748,"wave_count":1},"fallback_work":262861792,"features":{"max_depth":7,"packed_tokens":6310,"segment_count":19,"segments_below":[7,15,18,18,18,18,18,19]},"label":"minimum_effective_span_94","median_ms":910.8396911621094,"n":8,"spread_pct":2.457730657253581},{"cp_plan":{"max_rank_tokens":35328,"summary_ms":6.665421649813652,"wave_count":1},"fallback_work":2704285600,"features":{"max_depth":1,"packed_tokens":65842,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":4134.37353515625,"n":8,"spread_pct":0.4369686922366398},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.3294974863529205,"wave_count":1},"fallback_work":261373920,"features":{"max_depth":11,"packed_tokens":6245,"segment_count":23,"segments_below":[16,19,21,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":915.5038757324219,"n":8,"spread_pct":2.4188020019579834},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.302907899022102,"wave_count":1},"fallback_work":255020352,"features":{"max_depth":12,"packed_tokens":6083,"segment_count":24,"segments_below":[16,21,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":860.6584167480469,"n":8,"spread_pct":2.3744156232586597},{"cp_plan":{"max_rank_tokens":5632,"summary_ms":5.693880841135979,"wave_count":1},"fallback_work":250144224,"features":{"max_depth":13,"packed_tokens":5957,"segment_count":25,"segments_below":[17,24,24,24,24,24,24,25]},"label":"uniform_depth_12","median_ms":861.8241271972656,"n":8,"spread_pct":2.8394292359916182},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":5.854811519384384,"wave_count":1},"fallback_work":356853984,"features":{"max_depth":3,"packed_tokens":8629,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1008.18505859375,"n":8,"spread_pct":1.6791273381012737},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.454065442085266,"wave_count":1},"fallback_work":334987296,"features":{"max_depth":4,"packed_tokens":8089,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1010.1658325195312,"n":8,"spread_pct":1.3963095018493938},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.354719400405884,"wave_count":1},"fallback_work":320877088,"features":{"max_depth":5,"packed_tokens":7738,"segment_count":17,"segments_below":[8,9,11,14,16,16,16,17]},"label":"uniform_depth_4","median_ms":955.2257690429688,"n":8,"spread_pct":2.214950283439696},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.997704342007637,"wave_count":2},"fallback_work":310994080,"features":{"max_depth":6,"packed_tokens":7490,"segment_count":18,"segments_below":[10,11,13,15,17,17,17,18]},"label":"uniform_depth_5","median_ms":971.2817077636719,"n":8,"spread_pct":2.105667713965384},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.3684191554784775,"wave_count":2},"fallback_work":302383392,"features":{"max_depth":7,"packed_tokens":7273,"segment_count":19,"segments_below":[11,13,14,17,18,18,18,19]},"label":"uniform_depth_6","median_ms":972.2505187988281,"n":8,"spread_pct":1.1960435931090105},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.454698741436005,"wave_count":2},"fallback_work":290366304,"features":{"max_depth":8,"packed_tokens":6973,"segment_count":20,"segments_below":[14,14,15,19,19,19,19,20]},"label":"uniform_depth_7","median_ms":975.0990295410156,"n":8,"spread_pct":1.5548676148897511},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.382349878549576,"wave_count":2},"fallback_work":283274016,"features":{"max_depth":9,"packed_tokens":6793,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":928.2083129882812,"n":8,"spread_pct":2.337956958172355},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.22790290415287,"wave_count":2},"fallback_work":272406048,"features":{"max_depth":10,"packed_tokens":6521,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":929.2933959960938,"n":8,"spread_pct":3.242610132196927}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"8073af58c5c73ccd625f67dfee0685549b1e6fd8a81e602bd2dcb484d43c3e41","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":6,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.027642637491226,"wave_count":2},"fallback_work":506028352,"features":{"max_depth":2,"packed_tokens":12291,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1324.1172485351562,"n":8,"spread_pct":0.7503340618337403},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":9.474381804466248,"wave_count":1},"fallback_work":968631232,"features":{"max_depth":2,"packed_tokens":23563,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11339","median_ms":1962.0956420898438,"n":8,"spread_pct":0.5903821432654239},{"cp_plan":{"max_rank_tokens":20480,"summary_ms":6.683323532342911,"wave_count":1},"fallback_work":1433983712,"features":{"max_depth":2,"packed_tokens":34902,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11350","median_ms":2601.3521728515625,"n":8,"spread_pct":0.8508195365081364},{"cp_plan":{"max_rank_tokens":26624,"summary_ms":9.621523320674896,"wave_count":1},"fallback_work":1896053152,"features":{"max_depth":2,"packed_tokens":46161,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11352","median_ms":3385.8736572265625,"n":8,"spread_pct":0.6026878751499302},{"cp_plan":{"max_rank_tokens":31232,"summary_ms":7.152531296014786,"wave_count":1},"fallback_work":2356686112,"features":{"max_depth":2,"packed_tokens":57385,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11375","median_ms":3934.4412841796875,"n":8,"spread_pct":0.45344003741994143},{"cp_plan":{"max_rank_tokens":36864,"summary_ms":6.94306381046772,"wave_count":1},"fallback_work":2821669312,"features":{"max_depth":2,"packed_tokens":68715,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11425","median_ms":4673.3828125,"n":8,"spread_pct":0.6639788496354856},{"cp_plan":{"max_rank_tokens":43008,"summary_ms":7.826583459973335,"wave_count":1},"fallback_work":3292233952,"features":{"max_depth":2,"packed_tokens":80181,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11472","median_ms":5426.644775390625,"n":8,"spread_pct":0.3857017536400907},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":7.102804258465767,"wave_count":2},"fallback_work":490750880,"features":{"max_depth":6,"packed_tokens":11890,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":1255.3192749023438,"n":8,"spread_pct":1.2174948294977754},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.474952980875969,"wave_count":2},"fallback_work":500659744,"features":{"max_depth":3,"packed_tokens":12153,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1325.9755859375,"n":8,"spread_pct":0.6501790765790814},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":9.443433955311775,"wave_count":2},"fallback_work":490989568,"features":{"max_depth":5,"packed_tokens":11903,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1255.4515380859375,"n":8,"spread_pct":1.085937803090992},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":9.249968454241753,"wave_count":2},"fallback_work":490471232,"features":{"max_depth":7,"packed_tokens":11876,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":1256.1007080078125,"n":8,"spread_pct":60.19514912424002},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":7.085386663675308,"wave_count":2},"fallback_work":492172096,"features":{"max_depth":4,"packed_tokens":11939,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1328.9187622070312,"n":8,"spread_pct":0.7918612979696646},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":8.092369884252548,"wave_count":2},"fallback_work":493895776,"features":{"max_depth":4,"packed_tokens":11981,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1328.36572265625,"n":8,"spread_pct":0.7996712360590525},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":8.753083646297455,"wave_count":2},"fallback_work":495734944,"features":{"max_depth":3,"packed_tokens":12033,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1325.1085205078125,"n":8,"spread_pct":1.1236550219897417},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":9.358884766697884,"wave_count":2},"fallback_work":490381600,"features":{"max_depth":6,"packed_tokens":11881,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":1256.9048461914062,"n":8,"spread_pct":0.5910903583423384},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.5450815707445145,"wave_count":2},"fallback_work":495406624,"features":{"max_depth":3,"packed_tokens":12025,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1326.9683227539062,"n":8,"spread_pct":50.955563367710546},{"cp_plan":{"max_rank_tokens":49280,"summary_ms":7.539069280028343,"wave_count":1},"fallback_work":3762914080,"features":{"max_depth":1,"packed_tokens":91657,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":6212.011962890625,"n":8,"spread_pct":0.5536620807068714},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":7.487103343009949,"wave_count":2},"fallback_work":503614624,"features":{"max_depth":3,"packed_tokens":12225,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1324.5343627929688,"n":8,"spread_pct":2.144022550672496},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":8.202316239476204,"wave_count":2},"fallback_work":503499136,"features":{"max_depth":4,"packed_tokens":12215,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1325.60693359375,"n":8,"spread_pct":0.7966022344570202},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":7.781548425555229,"wave_count":2},"fallback_work":500018368,"features":{"max_depth":5,"packed_tokens":12123,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1324.5711669921875,"n":8,"spread_pct":0.9423653568597664},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":8.246401324868202,"wave_count":2},"fallback_work":494157280,"features":{"max_depth":6,"packed_tokens":11973,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":1329.6868896484375,"n":8,"spread_pct":2.058617645884275},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":9.368481114506721,"wave_count":2},"fallback_work":490594432,"features":{"max_depth":7,"packed_tokens":11879,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":1263.513671875,"n":8,"spread_pct":1.4006873348904576},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":9.726595133543015,"wave_count":2},"fallback_work":490684064,"features":{"max_depth":8,"packed_tokens":11874,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":1254.4848022460938,"n":8,"spread_pct":1.491679340514567}],"cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"8bb8a41e2c62e46d311f4fb98347ba7ee721963178c347ac157cccb0b8ec96a9","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":7,"kind":"ellavox-qwen3"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":74719,"summary_ms":0.00056438148021698,"wave_count":0},"fallback_work":3074477952,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":13112.80419921875,"n":8,"spread_pct":0.529770538643739},{"cp_plan":{"max_rank_tokens":14302,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":592293632,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":2593.6513671875,"n":8,"spread_pct":1.0984416199561804},{"cp_plan":{"max_rank_tokens":106432,"summary_ms":0.0013355165719985962,"wave_count":0},"fallback_work":4377272320,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":18700.93359375,"n":8,"spread_pct":0.7878116044925063},{"cp_plan":{"max_rank_tokens":13279,"summary_ms":0.0003129243850708008,"wave_count":0},"fallback_work":550489984,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":2430.645751953125,"n":8,"spread_pct":1.2816990887335933}],"cell":"cal-grpo-g16|Qwen/Qwen3-14B|L40|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":38879,"summary_ms":99.48470257222652,"wave_count":1},"fallback_work":3072005632,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":7568.21142578125,"n":8,"spread_pct":7.240573799603954},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":7.8962016850709915,"wave_count":1},"fallback_work":592786784,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1938.6018676757812,"n":8,"spread_pct":32.039843268317696},{"cp_plan":{"max_rank_tokens":55296,"summary_ms":10.915160179138184,"wave_count":1},"fallback_work":4373212160,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":10393.61962890625,"n":8,"spread_pct":0.7293856521520372},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":8.092327043414116,"wave_count":1},"fallback_work":551097856,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1787.540771484375,"n":8,"spread_pct":1.453972790994305}],"cell":"cal-grpo-g16|Qwen/Qwen3-14B|L40|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19456,"summary_ms":7.784722372889519,"wave_count":1},"fallback_work":3072440576,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":3812.9039306640625,"n":8,"spread_pct":0.690551836705057},{"cp_plan":{"max_rank_tokens":4608,"summary_ms":4.392847418785095,"wave_count":3},"fallback_work":594802944,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1065.9868774414062,"n":8,"spread_pct":3.226875678960058},{"cp_plan":{"max_rank_tokens":26880,"summary_ms":8.573267608880997,"wave_count":1},"fallback_work":4372754944,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":5115.1826171875,"n":8,"spread_pct":10.824450877414376},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":4.483306780457497,"wave_count":1},"fallback_work":553269504,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":908.0004577636719,"n":8,"spread_pct":4.548804052678557}],"cell":"cal-grpo-g16|Qwen/Qwen3-14B|L40|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":74719,"summary_ms":0.0002905726432800293,"wave_count":0},"fallback_work":3074477952,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":7866.4130859375,"n":8,"spread_pct":0.43828792978256176},{"cp_plan":{"max_rank_tokens":14302,"summary_ms":0.0002477318048477173,"wave_count":0},"fallback_work":592293632,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1601.38037109375,"n":8,"spread_pct":0.3777411272684363},{"cp_plan":{"max_rank_tokens":106432,"summary_ms":0.0003986060619354248,"wave_count":0},"fallback_work":4377272320,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":11082.150390625,"n":8,"spread_pct":0.2321617620959662},{"cp_plan":{"max_rank_tokens":13279,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":550489984,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1505.5513916015625,"n":8,"spread_pct":0.5330127100411344}],"cell":"cal-grpo-g16|Qwen/Qwen3-14B|L40|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":38879,"summary_ms":8.0448929220438,"wave_count":1},"fallback_work":3072005632,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":4476.441650390625,"n":8,"spread_pct":0.30440394196941944},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":7.68100842833519,"wave_count":1},"fallback_work":592786784,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1199.33984375,"n":8,"spread_pct":44.18603088938902},{"cp_plan":{"max_rank_tokens":55296,"summary_ms":7.27500393986702,"wave_count":1},"fallback_work":4373212160,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":6078.711669921875,"n":8,"spread_pct":0.6965105352339265},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":7.8023746609687805,"wave_count":1},"fallback_work":551097856,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1128.7276611328125,"n":8,"spread_pct":54.08236618747105}],"cell":"cal-grpo-g16|Qwen/Qwen3-14B|L40|tp2|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":58351,"summary_ms":0.0003390014171600342,"wave_count":0},"fallback_work":2399852416,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":10460.34814453125,"n":8,"spread_pct":0.6813957451766444},{"cp_plan":{"max_rank_tokens":24556,"summary_ms":0.00048242509365081787,"wave_count":0},"fallback_work":1011971584,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":4412.23486328125,"n":8,"spread_pct":0.8857428468446615},{"cp_plan":{"max_rank_tokens":73696,"summary_ms":0.001130625605583191,"wave_count":0},"fallback_work":3030118400,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":12911.478515625,"n":8,"spread_pct":0.651173851029176},{"cp_plan":{"max_rank_tokens":21487,"summary_ms":0.0002775341272354126,"wave_count":0},"fallback_work":886101888,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":3964.1087646484375,"n":8,"spread_pct":0.8332948666376855}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":31727,"summary_ms":10.698914527893066,"wave_count":1},"fallback_work":2397641472,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":6339.59375,"n":8,"spread_pct":0.6955989022364406},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":8.631149306893349,"wave_count":1},"fallback_work":1011579328,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":2940.083984375,"n":8,"spread_pct":9.326304055674429},{"cp_plan":{"max_rank_tokens":37888,"summary_ms":6.4080096781253815,"wave_count":1},"fallback_work":3027105280,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":7283.373046875,"n":8,"spread_pct":2.581402632869654},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":10.3529691696167,"wave_count":1},"fallback_work":885922560,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":2790.142822265625,"n":8,"spread_pct":0.9892876752483392}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":15360,"summary_ms":7.314587011933327,"wave_count":1},"fallback_work":2397420672,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3063.466552734375,"n":8,"spread_pct":18.705958474999797},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.6147312968969345,"wave_count":1},"fallback_work":1012562848,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1362.6600341796875,"n":8,"spread_pct":46.7867297338481},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":21.91111631691456,"wave_count":1},"fallback_work":3026385152,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":3645.3988037109375,"n":8,"spread_pct":0.4985482588934902},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.526291042566299,"wave_count":1},"fallback_work":887110784,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1263.3612670898438,"n":8,"spread_pct":0.8469141202645879}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":58351,"summary_ms":0.0004600733518600464,"wave_count":0},"fallback_work":2399852416,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":6245.81689453125,"n":8,"spread_pct":0.4694001055237517},{"cp_plan":{"max_rank_tokens":24556,"summary_ms":0.00018812716007232666,"wave_count":0},"fallback_work":1011971584,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":2654.2012939453125,"n":8,"spread_pct":0.9559486988516925},{"cp_plan":{"max_rank_tokens":73696,"summary_ms":0.0003334134817123413,"wave_count":0},"fallback_work":3030118400,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":7687.751953125,"n":8,"spread_pct":0.3356279735750531},{"cp_plan":{"max_rank_tokens":21487,"summary_ms":0.00021420419216156006,"wave_count":0},"fallback_work":886101888,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":2408.3172607421875,"n":8,"spread_pct":0.4478903701593624}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":31727,"summary_ms":11.245757341384888,"wave_count":1},"fallback_work":2397641472,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3753.244384765625,"n":8,"spread_pct":0.43930742471568707},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":8.59602726995945,"wave_count":1},"fallback_work":1011579328,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1746.8931884765625,"n":8,"spread_pct":36.48899144260643},{"cp_plan":{"max_rank_tokens":37888,"summary_ms":7.177593186497688,"wave_count":1},"fallback_work":3027105280,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":4261.24755859375,"n":8,"spread_pct":0.4290460620359607},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":10.160276666283607,"wave_count":1},"fallback_work":885922560,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1659.724609375,"n":8,"spread_pct":0.44799883263766227}],"cell":"cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp2|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":37871,"summary_ms":0.0009145587682723999,"wave_count":0},"fallback_work":1558370176,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":6609.2001953125,"n":8,"spread_pct":0.216546802877913},{"cp_plan":{"max_rank_tokens":10222,"summary_ms":0.00034458935260772705,"wave_count":0},"fallback_work":422557440,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1798.0990600585938,"n":8,"spread_pct":0.3250842539160229},{"cp_plan":{"max_rank_tokens":53216,"summary_ms":0.0011883676052093506,"wave_count":0},"fallback_work":2188636160,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":9037.677734375,"n":8,"spread_pct":0.45089016446123664},{"cp_plan":{"max_rank_tokens":9199,"summary_ms":0.0005736947059631348,"wave_count":0},"fallback_work":380753792,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1650.2720336914062,"n":8,"spread_pct":0.5341138233416699}],"cell":"cal-grpo-g8|Qwen/Qwen3-14B|L40|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":21487,"summary_ms":9.30088758468628,"wave_count":1},"fallback_work":1557142272,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":4262.13427734375,"n":8,"spread_pct":0.5398648688126306},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.326822564005852,"wave_count":1},"fallback_work":422722144,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1411.1672973632812,"n":8,"spread_pct":1.1562952721959865},{"cp_plan":{"max_rank_tokens":28672,"summary_ms":10.628854855895042,"wave_count":1},"fallback_work":2186606080,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":5397.746337890625,"n":8,"spread_pct":0.39201835220491554},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":8.035432547330856,"wave_count":1},"fallback_work":381033216,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1232.760009765625,"n":8,"spread_pct":1.6694706385035012}],"cell":"cal-grpo-g8|Qwen/Qwen3-14B|L40|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.760444328188896,"wave_count":1},"fallback_work":1557412992,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":2017.8770751953125,"n":8,"spread_pct":23.880476567817364},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.763219341635704,"wave_count":3},"fallback_work":423787648,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":886.9709167480469,"n":8,"spread_pct":1.543620768914732},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":18.76411959528923,"wave_count":1},"fallback_work":2186377472,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":2581.4610595703125,"n":8,"spread_pct":23.509266820410275},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.883577883243561,"wave_count":1},"fallback_work":382254208,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":735.5798645019531,"n":8,"spread_pct":8.481928398640843}],"cell":"cal-grpo-g8|Qwen/Qwen3-14B|L40|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":37871,"summary_ms":0.00042654573917388916,"wave_count":0},"fallback_work":1558370176,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3990.1629638671875,"n":8,"spread_pct":0.20458026430049606},{"cp_plan":{"max_rank_tokens":10222,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":422557440,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1120.4286499023438,"n":8,"spread_pct":0.689030289801174},{"cp_plan":{"max_rank_tokens":53216,"summary_ms":0.0006835907697677612,"wave_count":0},"fallback_work":2188636160,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":5459.180908203125,"n":8,"spread_pct":0.22418691992528889},{"cp_plan":{"max_rank_tokens":9199,"summary_ms":0.0003390014171600342,"wave_count":0},"fallback_work":380753792,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1029.4843139648438,"n":8,"spread_pct":0.826972270407826}],"cell":"cal-grpo-g8|Qwen/Qwen3-14B|L40|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":21487,"summary_ms":8.921809494495392,"wave_count":1},"fallback_work":1557142272,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":2527.6839599609375,"n":8,"spread_pct":1.1575142005372443},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.237851455807686,"wave_count":1},"fallback_work":422722144,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":906.0047302246094,"n":8,"spread_pct":0.9376594817232828},{"cp_plan":{"max_rank_tokens":28672,"summary_ms":9.267108514904976,"wave_count":1},"fallback_work":2186606080,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":3167.3299560546875,"n":8,"spread_pct":0.6073511971645639},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.79452919960022,"wave_count":1},"fallback_work":381033216,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":832.88720703125,"n":8,"spread_pct":66.96376908028302}],"cell":"cal-grpo-g8|Qwen/Qwen3-14B|L40|tp2|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":28852,"summary_ms":0.00030919909477233887,"wave_count":0},"fallback_work":1188485632,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":5129.181396484375,"n":8,"spread_pct":0.36456536247317656},{"cp_plan":{"max_rank_tokens":33460,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":1377360384,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":5837.13134765625,"n":8,"spread_pct":0.13160795001357237},{"cp_plan":{"max_rank_tokens":29364,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":1209293312,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":5186.228271484375,"n":8,"spread_pct":0.35354111425377205},{"cp_plan":{"max_rank_tokens":36532,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":1503353344,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":6396.285888671875,"n":8,"spread_pct":0.1176372381326184},{"cp_plan":{"max_rank_tokens":48820,"summary_ms":0.0008400529623031616,"wave_count":0},"fallback_work":2008013312,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":8735.63671875,"n":8,"spread_pct":0.1880654091560119}],"cell":"cal-hetero2|Qwen/Qwen3-14B|L40|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":18432,"summary_ms":5.537401884794235,"wave_count":1},"fallback_work":1187887168,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":3551.2147216796875,"n":8,"spread_pct":0.5571445812565993},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":4.0008556097745895,"wave_count":1},"fallback_work":1376409664,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":4269.972412109375,"n":8,"spread_pct":0.353794641644702},{"cp_plan":{"max_rank_tokens":18944,"summary_ms":5.46645000576973,"wave_count":1},"fallback_work":1208604736,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":3629.0909423828125,"n":8,"spread_pct":0.49847436115082006},{"cp_plan":{"max_rank_tokens":25088,"summary_ms":5.020478740334511,"wave_count":1},"fallback_work":1502189632,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":4718.543701171875,"n":8,"spread_pct":0.3405157148064472},{"cp_plan":{"max_rank_tokens":30720,"summary_ms":5.117565393447876,"wave_count":1},"fallback_work":2006194240,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":5815.84375,"n":8,"spread_pct":9.003061746737092}],"cell":"cal-hetero2|Qwen/Qwen3-14B|L40|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13312,"summary_ms":12.050503864884377,"wave_count":3},"fallback_work":1188767584,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":2545.5723876953125,"n":8,"spread_pct":0.7953837969947148},{"cp_plan":{"max_rank_tokens":16384,"summary_ms":5.199875682592392,"wave_count":2},"fallback_work":1376917344,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":3069.3839111328125,"n":8,"spread_pct":0.5804077274802378},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":12.189352884888649,"wave_count":3},"fallback_work":1209341792,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":2641.642578125,"n":8,"spread_pct":0.7613282899119496},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":5.554156377911568,"wave_count":2},"fallback_work":1502492512,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":3321.7982177734375,"n":8,"spread_pct":0.6057877557803087},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":6.307011470198631,"wave_count":1},"fallback_work":2006071136,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":4210.11328125,"n":8,"spread_pct":0.6133853657812428}],"cell":"cal-hetero2|Qwen/Qwen3-14B|L40|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":28852,"summary_ms":0.0002998858690261841,"wave_count":0},"fallback_work":1188485632,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":3075.08935546875,"n":8,"spread_pct":0.5243114918376832},{"cp_plan":{"max_rank_tokens":33460,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":1377360384,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":3537.135498046875,"n":8,"spread_pct":0.31864066161300436},{"cp_plan":{"max_rank_tokens":29364,"summary_ms":0.0002812594175338745,"wave_count":0},"fallback_work":1209293312,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":3123.227783203125,"n":8,"spread_pct":0.30132713544345185},{"cp_plan":{"max_rank_tokens":36532,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":1503353344,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":3847.27685546875,"n":8,"spread_pct":0.31536739092166827},{"cp_plan":{"max_rank_tokens":48820,"summary_ms":0.00029616057872772217,"wave_count":0},"fallback_work":2008013312,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":5233.918701171875,"n":8,"spread_pct":0.35017045095955424}],"cell":"cal-hetero2|Qwen/Qwen3-14B|L40|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":18432,"summary_ms":5.499131977558136,"wave_count":1},"fallback_work":1187887168,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":2112.697265625,"n":8,"spread_pct":30.727090942876078},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":4.81068529188633,"wave_count":1},"fallback_work":1376409664,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":2519.66748046875,"n":8,"spread_pct":0.5457263066590003},{"cp_plan":{"max_rank_tokens":18944,"summary_ms":5.376774817705154,"wave_count":1},"fallback_work":1208604736,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":2155.22998046875,"n":8,"spread_pct":0.5621884990431388},{"cp_plan":{"max_rank_tokens":25088,"summary_ms":6.260095164179802,"wave_count":1},"fallback_work":1502189632,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":2778.874267578125,"n":8,"spread_pct":0.57827661602445},{"cp_plan":{"max_rank_tokens":30720,"summary_ms":5.737705156207085,"wave_count":1},"fallback_work":2006194240,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":3428.8206787109375,"n":8,"spread_pct":0.7058090813741045}],"cell":"cal-hetero2|Qwen/Qwen3-14B|L40|tp2|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16756,"summary_ms":0.0002477318048477173,"wave_count":0},"fallback_work":690371072,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":2934.2960205078125,"n":8,"spread_pct":0.6042578106198195},{"cp_plan":{"max_rank_tokens":17524,"summary_ms":0.00030919909477233887,"wave_count":0},"fallback_work":721697280,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":3032.2589111328125,"n":8,"spread_pct":0.3970972129323097},{"cp_plan":{"max_rank_tokens":25204,"summary_ms":0.0008996576070785522,"wave_count":0},"fallback_work":1037023744,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":4338.87353515625,"n":8,"spread_pct":0.23571876385956222}],"cell":"cal-hetero3|Qwen/Qwen3-14B|L40|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9588,"summary_ms":5.1960088312625885,"wave_count":1},"fallback_work":690058304,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1803.7870483398438,"n":8,"spread_pct":0.6462910826685135},{"cp_plan":{"max_rank_tokens":10356,"summary_ms":5.345230922102928,"wave_count":1},"fallback_work":721282112,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1929.6567993164062,"n":8,"spread_pct":0.47150927281010324},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":6.356490775942802,"wave_count":1},"fallback_work":1036174400,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":2660.6103515625,"n":8,"spread_pct":0.6848225247892705}],"cell":"cal-hetero3|Qwen/Qwen3-14B|L40|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":9.952997788786888,"wave_count":3},"fallback_work":690639200,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1187.2612915039062,"n":8,"spread_pct":2.4109801266074813},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":9.700940921902657,"wave_count":3},"fallback_work":721713504,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1261.9550170898438,"n":8,"spread_pct":1.244039489315831},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":4.452075809240341,"wave_count":3},"fallback_work":1036290400,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1385.114501953125,"n":8,"spread_pct":44.574198425678716}],"cell":"cal-hetero3|Qwen/Qwen3-14B|L40|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16756,"summary_ms":0.00021420419216156006,"wave_count":0},"fallback_work":690371072,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1786.3325805664062,"n":8,"spread_pct":0.6454923217479128},{"cp_plan":{"max_rank_tokens":17524,"summary_ms":0.0004973262548446655,"wave_count":0},"fallback_work":721697280,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1844.7918701171875,"n":8,"spread_pct":1.113532574118108},{"cp_plan":{"max_rank_tokens":25204,"summary_ms":0.00034086406230926514,"wave_count":0},"fallback_work":1037023744,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":2629.09619140625,"n":8,"spread_pct":0.26110669195706326}],"cell":"cal-hetero3|Qwen/Qwen3-14B|L40|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9588,"summary_ms":5.0531793385744095,"wave_count":1},"fallback_work":690058304,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1130.4411010742188,"n":8,"spread_pct":1.4669069901556313},{"cp_plan":{"max_rank_tokens":10356,"summary_ms":5.129674449563026,"wave_count":1},"fallback_work":721282112,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1198.5192260742188,"n":8,"spread_pct":0.8379684424021973},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":6.3176993280649185,"wave_count":1},"fallback_work":1036174400,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1593.5028686523438,"n":8,"spread_pct":0.589988850522969}],"cell":"cal-hetero3|Qwen/Qwen3-14B|L40|tp2|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19380,"summary_ms":0.00038929283618927,"wave_count":0},"fallback_work":798546432,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":3400.3017578125,"n":8,"spread_pct":0.31932195815571285},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":1029495296,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":4288.28369140625,"n":8,"spread_pct":0.15528776001562214},{"cp_plan":{"max_rank_tokens":20916,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":861428224,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":3615.9708251953125,"n":8,"spread_pct":0.4431768744569287},{"cp_plan":{"max_rank_tokens":29108,"summary_ms":0.001041218638420105,"wave_count":0},"fallback_work":1197562368,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":5057.95654296875,"n":8,"spread_pct":0.32359288501109457}],"cell":"cal-hetero|Qwen/Qwen3-14B|L40|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.801582872867584,"wave_count":1},"fallback_work":798206016,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":2165.6829833984375,"n":8,"spread_pct":0.7109071924086628},{"cp_plan":{"max_rank_tokens":15872,"summary_ms":5.575787276029587,"wave_count":1},"fallback_work":1028753472,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":2959.44775390625,"n":8,"spread_pct":0.6341251180301962},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":7.527470588684082,"wave_count":1},"fallback_work":860948544,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":2434.90869140625,"n":8,"spread_pct":1.4693505294992093},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":7.609955966472626,"wave_count":1},"fallback_work":1196558400,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":3272.9256591796875,"n":8,"spread_pct":0.6624618773363541}],"cell":"cal-hetero|Qwen/Qwen3-14B|L40|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6580,"summary_ms":4.150141030550003,"wave_count":3},"fallback_work":798920544,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1326.8385009765625,"n":8,"spread_pct":42.201062223784604},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":4.352135583758354,"wave_count":2},"fallback_work":1029070688,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1838.009033203125,"n":8,"spread_pct":1.1797061427365534},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.143113270401955,"wave_count":1},"fallback_work":861495136,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1468.2855834960938,"n":8,"spread_pct":50.538079465036965},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":9.5951147377491,"wave_count":3},"fallback_work":1196646240,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":2023.564453125,"n":8,"spread_pct":0.8419114630838798}],"cell":"cal-hetero|Qwen/Qwen3-14B|L40|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":1.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,4,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19380,"summary_ms":0.0002849847078323364,"wave_count":0},"fallback_work":798546432,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":2059.1854248046875,"n":8,"spread_pct":0.41150411761989425},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.00030919909477233887,"wave_count":0},"fallback_work":1029495296,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":2593.5263671875,"n":8,"spread_pct":0.5030177677216898},{"cp_plan":{"max_rank_tokens":20916,"summary_ms":0.00021792948246002197,"wave_count":0},"fallback_work":861428224,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":2186.82958984375,"n":8,"spread_pct":0.7632473077128326},{"cp_plan":{"max_rank_tokens":29108,"summary_ms":0.0003241002559661865,"wave_count":0},"fallback_work":1197562368,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":3059.332275390625,"n":8,"spread_pct":0.49329564022097155}],"cell":"cal-hetero|Qwen/Qwen3-14B|L40|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[2,1,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.693830713629723,"wave_count":1},"fallback_work":798206016,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1307.4312133789062,"n":8,"spread_pct":0.8791578974100732},{"cp_plan":{"max_rank_tokens":15872,"summary_ms":5.567241460084915,"wave_count":1},"fallback_work":1028753472,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1769.5880737304688,"n":8,"spread_pct":0.994304439244173},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":7.443800568580627,"wave_count":1},"fallback_work":860948544,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1467.5924072265625,"n":8,"spread_pct":36.352595716853635},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":7.489180192351341,"wave_count":1},"fallback_work":1196558400,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1950.2268676757812,"n":8,"spread_pct":0.6051166485569822}],"cell":"cal-hetero|Qwen/Qwen3-14B|L40|tp2|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":0.0,"layers":40.0,"tp":2.0,"uses_gdn":0.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":0,"gdn_key_head_dim":0,"gdn_key_heads":0,"gdn_value_head_dim":0,"gdn_value_heads":0,"hidden_size":5120,"kv_channels":128,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":40,"num_query_groups":8},"hidden_size":5120,"model":"Qwen/Qwen3-14B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[2,2,1,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"kind":"cal-hetero","seed":7777}} + ], + "fit_arguments": { + "exclude_cells": [], + "holdout": "", + "incumbent": "depth_one", + "objective": "regret", + "require_complete": 8, + "rerank_shapes": ["tp1cp4"], + "shortlist_size": 3, + "terms": ["token_per_rank","token_cp_exchange","token_tp_collective","gdn_token_per_rank","attention_token_cp_exchange","tiny_segment_per_layer","level_cp_per_layer","level_tp_per_layer","gdn_level","gdn_level_tp"] + }, + "integer_table_milli_us": {"attention_token_cp_exchange":0,"gdn_level":0,"gdn_level_tp":0,"gdn_token_per_rank":0,"level_cp_per_layer":0,"level_tp_per_layer":0,"tiny_segment_per_layer":0,"token_cp_exchange":0,"token_per_rank":4607,"token_tp_collective":487}, + "integer_table_sha256": "d053b7d4c1ff454ebedf29e5de50d89d32d1275c554c7708fb83f88b88ef3e64", + "manifest": { + "excluded": [], + "expected": ["cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g0","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g1","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g2","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g3","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g4","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g5","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g6","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g7","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g7","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g0","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g1","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g2","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g3","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g4","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g5","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g6","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g7","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g0","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g1","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g2","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g3","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g4","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g5","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g6","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g7","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g7","cal-grpo-g16|Qwen/Qwen3-14B|L40|tp1|cp1","cal-grpo-g16|Qwen/Qwen3-14B|L40|tp1|cp2","cal-grpo-g16|Qwen/Qwen3-14B|L40|tp1|cp4","cal-grpo-g16|Qwen/Qwen3-14B|L40|tp2|cp1","cal-grpo-g16|Qwen/Qwen3-14B|L40|tp2|cp2","cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp1|cp1","cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp1|cp2","cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp1|cp4","cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp2|cp1","cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp2|cp2","cal-grpo-g8|Qwen/Qwen3-14B|L40|tp1|cp1","cal-grpo-g8|Qwen/Qwen3-14B|L40|tp1|cp2","cal-grpo-g8|Qwen/Qwen3-14B|L40|tp1|cp4","cal-grpo-g8|Qwen/Qwen3-14B|L40|tp2|cp1","cal-grpo-g8|Qwen/Qwen3-14B|L40|tp2|cp2","cal-hetero2|Qwen/Qwen3-14B|L40|tp1|cp1","cal-hetero2|Qwen/Qwen3-14B|L40|tp1|cp2","cal-hetero2|Qwen/Qwen3-14B|L40|tp1|cp4","cal-hetero2|Qwen/Qwen3-14B|L40|tp2|cp1","cal-hetero2|Qwen/Qwen3-14B|L40|tp2|cp2","cal-hetero3|Qwen/Qwen3-14B|L40|tp1|cp1","cal-hetero3|Qwen/Qwen3-14B|L40|tp1|cp2","cal-hetero3|Qwen/Qwen3-14B|L40|tp1|cp4","cal-hetero3|Qwen/Qwen3-14B|L40|tp2|cp1","cal-hetero3|Qwen/Qwen3-14B|L40|tp2|cp2","cal-hetero|Qwen/Qwen3-14B|L40|tp1|cp1","cal-hetero|Qwen/Qwen3-14B|L40|tp1|cp2","cal-hetero|Qwen/Qwen3-14B|L40|tp1|cp4","cal-hetero|Qwen/Qwen3-14B|L40|tp2|cp1","cal-hetero|Qwen/Qwen3-14B|L40|tp2|cp2"], + "path": "trainer_rank_cost_calibration_manifest_dense-attn-h5120-h200-bf16.json" + }, + "measured_envelope": {"device_names":["NVIDIA H200"],"hidden_sizes":[5120],"param_dtypes":["torch.bfloat16"]}, + "metrics": { + "integer_all": {"cells":56,"clear_misses":[],"max_regret_pct":0.9859102120967009,"median_regret_pct":0.0,"ordered_pairs":4550,"p95_regret_pct":0.4984641043130954,"pairwise_accuracy":0.9991208791208791}, + "test": {"cells":16,"clear_misses":[],"max_regret_pct":0.4984641043130954,"median_regret_pct":0.0,"ordered_pairs":931,"p95_regret_pct":0.4984641043130954,"pairwise_accuracy":0.9978517722878625}, + "train": {"cells":40,"clear_misses":[],"max_regret_pct":0.9859102120967009,"median_regret_pct":0.0,"ordered_pairs":3619,"p95_regret_pct":0.5450747471522305,"pairwise_accuracy":0.9994473611494888} + }, + "reranker": {"incumbent":"depth_one","max_rank_token_per_layer_milli_us":4642,"metrics":{"all":{"cells":14,"clear_misses":[],"max_regret_pct":4.131367479106448,"median_regret_pct":0.0,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":1.2154640278075564,"savings_vs_fallback_mean_pct":1.1855217379426453,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":34,"p95_regret_pct":4.131367479106448,"pairwise_accuracy":0.9705882352941176,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g0":{"best":"depth_one","best_lead_pct":0.027239369197383734,"cheap_regret_pct":0.17468259630448205,"fallback_regret_pct":0.027239369197383734,"irrecoverable_pct":0.0,"planning_ms":12.589329853653908,"regret_pct":0.17468259630448205,"selected":"uniform_depth_3","shortlist":["uniform_depth_3","uniform_depth_2","minimum_effective_span_46","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g1":{"best":"depth_one","best_lead_pct":42.07464661437225,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.695261552929878,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g2":{"best":"no_sharing","best_lead_pct":0.0,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":0.0,"regret_pct":0.0,"selected":"no_sharing","shortlist":["no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g3":{"best":"minimum_effective_span_278","best_lead_pct":8.93687578804834,"cheap_regret_pct":17.01649638930579,"fallback_regret_pct":17.01649638930579,"irrecoverable_pct":0.0,"planning_ms":22.446637973189354,"regret_pct":0.0,"selected":"minimum_effective_span_278","shortlist":["uniform_depth_3","minimum_effective_span_278","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g4":{"best":"uniform_depth_2","best_lead_pct":16.24776157838761,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.5343706458807,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_6717"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g5":{"best":"minimum_effective_span_22","best_lead_pct":0.23565074352816617,"cheap_regret_pct":0.23565074352816617,"fallback_regret_pct":0.2847419415074594,"irrecoverable_pct":0.0,"planning_ms":14.467140659689903,"regret_pct":0.23565074352816617,"selected":"uniform_depth_4","shortlist":["uniform_depth_4","uniform_depth_3","minimum_effective_span_22","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g6":{"best":"minimum_effective_span_71","best_lead_pct":0.07292714639813859,"cheap_regret_pct":4.131367479106448,"fallback_regret_pct":4.131367479106448,"irrecoverable_pct":3.500619830685709,"planning_ms":11.686466634273529,"regret_pct":4.131367479106448,"selected":"uniform_depth_12","shortlist":["uniform_depth_12","minimum_effective_span_32","minimum_effective_span_37","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g7":{"best":"minimum_effective_span_48","best_lead_pct":0.11853738993800103,"cheap_regret_pct":0.4393774189189527,"fallback_regret_pct":0.11853738993800103,"irrecoverable_pct":0.16261266446842892,"planning_ms":12.027047574520111,"regret_pct":0.4393774189189527,"selected":"uniform_depth_7","shortlist":["uniform_depth_7","minimum_effective_span_3","uniform_depth_6","depth_one"]},"cal-grpo-g16|Qwen/Qwen3-14B|L40|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":17.399376655254287,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.268029153347015,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":7.8598869283509645,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":11.929318308830261,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g8|Qwen/Qwen3-14B|L40|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":20.58118493341273,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":9.644022211432457,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-hetero2|Qwen/Qwen3-14B|L40|tp1|cp4":{"best":"depth_one","best_lead_pct":3.7740113341136086,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":24.239856749773026,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_257","minimum_effective_span_1025"]},"cal-hetero3|Qwen/Qwen3-14B|L40|tp1|cp4":{"best":"depth_one","best_lead_pct":6.291262599096684,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":19.653938710689545,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_769","no_sharing"]},"cal-hetero|Qwen/Qwen3-14B|L40|tp1|cp4":{"best":"depth_one","best_lead_pct":10.660459612486765,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.502276614308357,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_513","minimum_effective_span_2049"]}},"planning_cost":{"max_cell":"cal-hetero3|Qwen/Qwen3-14B|L40|tp1|cp4","max_ms":19.653938710689545,"max_pct_of_cell":1.6554012879333293,"mean_pct_of_cell":1.0451049086024087,"median_pct_of_cell":0.9617408107450203},"recall":{"best_in_shortlist":12,"cells":14,"clear_winner_cells":7,"clear_winners_in_shortlist":7,"max_irrecoverable_pct":3.500619830685709}},"by_shape":{"tp1cp4":{"cells":14,"clear_misses":[],"max_regret_pct":4.131367479106448,"median_regret_pct":0.0,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":1.2154640278075564,"savings_vs_fallback_mean_pct":1.1855217379426453,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":34,"p95_regret_pct":4.131367479106448,"pairwise_accuracy":0.9705882352941176,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g0":{"best":"depth_one","best_lead_pct":0.027239369197383734,"cheap_regret_pct":0.17468259630448205,"fallback_regret_pct":0.027239369197383734,"irrecoverable_pct":0.0,"planning_ms":12.589329853653908,"regret_pct":0.17468259630448205,"selected":"uniform_depth_3","shortlist":["uniform_depth_3","uniform_depth_2","minimum_effective_span_46","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g1":{"best":"depth_one","best_lead_pct":42.07464661437225,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.695261552929878,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g2":{"best":"no_sharing","best_lead_pct":0.0,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":0.0,"regret_pct":0.0,"selected":"no_sharing","shortlist":["no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g3":{"best":"minimum_effective_span_278","best_lead_pct":8.93687578804834,"cheap_regret_pct":17.01649638930579,"fallback_regret_pct":17.01649638930579,"irrecoverable_pct":0.0,"planning_ms":22.446637973189354,"regret_pct":0.0,"selected":"minimum_effective_span_278","shortlist":["uniform_depth_3","minimum_effective_span_278","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g4":{"best":"uniform_depth_2","best_lead_pct":16.24776157838761,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.5343706458807,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","depth_one","minimum_effective_span_6717"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g5":{"best":"minimum_effective_span_22","best_lead_pct":0.23565074352816617,"cheap_regret_pct":0.23565074352816617,"fallback_regret_pct":0.2847419415074594,"irrecoverable_pct":0.0,"planning_ms":14.467140659689903,"regret_pct":0.23565074352816617,"selected":"uniform_depth_4","shortlist":["uniform_depth_4","uniform_depth_3","minimum_effective_span_22","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g6":{"best":"minimum_effective_span_71","best_lead_pct":0.07292714639813859,"cheap_regret_pct":4.131367479106448,"fallback_regret_pct":4.131367479106448,"irrecoverable_pct":3.500619830685709,"planning_ms":11.686466634273529,"regret_pct":4.131367479106448,"selected":"uniform_depth_12","shortlist":["uniform_depth_12","minimum_effective_span_32","minimum_effective_span_37","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g7":{"best":"minimum_effective_span_48","best_lead_pct":0.11853738993800103,"cheap_regret_pct":0.4393774189189527,"fallback_regret_pct":0.11853738993800103,"irrecoverable_pct":0.16261266446842892,"planning_ms":12.027047574520111,"regret_pct":0.4393774189189527,"selected":"uniform_depth_7","shortlist":["uniform_depth_7","minimum_effective_span_3","uniform_depth_6","depth_one"]},"cal-grpo-g16|Qwen/Qwen3-14B|L40|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":17.399376655254287,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":12.268029153347015,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":7.8598869283509645,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":11.929318308830261,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-grpo-g8|Qwen/Qwen3-14B|L40|tp1|cp4":{"best":"uniform_depth_2","best_lead_pct":20.58118493341273,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":9.644022211432457,"regret_pct":0.0,"selected":"uniform_depth_2","shortlist":["uniform_depth_2","minimum_effective_span_1024","depth_one"]},"cal-hetero2|Qwen/Qwen3-14B|L40|tp1|cp4":{"best":"depth_one","best_lead_pct":3.7740113341136086,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":24.239856749773026,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_257","minimum_effective_span_1025"]},"cal-hetero3|Qwen/Qwen3-14B|L40|tp1|cp4":{"best":"depth_one","best_lead_pct":6.291262599096684,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":19.653938710689545,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_769","no_sharing"]},"cal-hetero|Qwen/Qwen3-14B|L40|tp1|cp4":{"best":"depth_one","best_lead_pct":10.660459612486765,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.502276614308357,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","minimum_effective_span_513","minimum_effective_span_2049"]}},"planning_cost":{"max_cell":"cal-hetero3|Qwen/Qwen3-14B|L40|tp1|cp4","max_ms":19.653938710689545,"max_pct_of_cell":1.6554012879333293,"mean_pct_of_cell":1.0451049086024087,"median_pct_of_cell":0.9617408107450203},"recall":{"best_in_shortlist":12,"cells":14,"clear_winner_cells":7,"clear_winners_in_shortlist":7,"max_irrecoverable_pct":3.500619830685709}}},"held_out":{"cells":4,"clear_misses":[],"max_regret_pct":0.4393774189189527,"median_regret_pct":0.11782537176408309,"non_regression":{"fallback_available":true,"savings_vs_cheap_mean_pct":4.254124097326447,"savings_vs_fallback_mean_pct":4.186186889576033,"worse_than_cheap_cells":[],"worse_than_fallback_cells":[]},"ordered_pairs":10,"p95_regret_pct":0.4393774189189527,"pairwise_accuracy":1.0,"per_cell":{"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g1":{"best":"depth_one","best_lead_pct":42.07464661437225,"cheap_regret_pct":0.0,"fallback_regret_pct":0.0,"irrecoverable_pct":0.0,"planning_ms":8.695261552929878,"regret_pct":0.0,"selected":"depth_one","shortlist":["depth_one","no_sharing"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g3":{"best":"minimum_effective_span_278","best_lead_pct":8.93687578804834,"cheap_regret_pct":17.01649638930579,"fallback_regret_pct":17.01649638930579,"irrecoverable_pct":0.0,"planning_ms":22.446637973189354,"regret_pct":0.0,"selected":"minimum_effective_span_278","shortlist":["uniform_depth_3","minimum_effective_span_278","uniform_depth_2","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g5":{"best":"minimum_effective_span_22","best_lead_pct":0.23565074352816617,"cheap_regret_pct":0.23565074352816617,"fallback_regret_pct":0.2847419415074594,"irrecoverable_pct":0.0,"planning_ms":14.467140659689903,"regret_pct":0.23565074352816617,"selected":"uniform_depth_4","shortlist":["uniform_depth_4","uniform_depth_3","minimum_effective_span_22","depth_one"]},"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g7":{"best":"minimum_effective_span_48","best_lead_pct":0.11853738993800103,"cheap_regret_pct":0.4393774189189527,"fallback_regret_pct":0.11853738993800103,"irrecoverable_pct":0.16261266446842892,"planning_ms":12.027047574520111,"regret_pct":0.4393774189189527,"selected":"uniform_depth_7","shortlist":["uniform_depth_7","minimum_effective_span_3","uniform_depth_6","depth_one"]}},"planning_cost":{"max_cell":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g3","max_ms":22.446637973189354,"max_pct_of_cell":1.5758962097737144,"mean_pct_of_cell":1.0647475870147265,"median_pct_of_cell":1.099921412172461},"recall":{"best_in_shortlist":3,"cells":4,"clear_winner_cells":1,"clear_winners_in_shortlist":1,"max_irrecoverable_pct":0.16261266446842892}}},"shapes":["tp1cp4"],"shortlist_size":3,"shortlist_table_milli_us":{"attention_token_cp_exchange":0,"gdn_level":0,"gdn_level_tp":0,"gdn_token_per_rank":0,"level_cp_per_layer":0,"level_tp_per_layer":0,"tiny_segment_per_layer":0,"token_cp_exchange":0,"token_per_rank":4592,"token_tp_collective":496},"wave_per_layer_milli_us":654967}, + "schema": "art.dev.trainer_rank_cost_calibration_certificate.v1", + "table_id": "dense-attn-h5120-h200-bf16" +} diff --git a/dev/trainer_rank_cost_calibration_certificate_dense-gdn-h5120-h200-bf16.json b/dev/trainer_rank_cost_calibration_certificate_dense-gdn-h5120-h200-bf16.json new file mode 100644 index 000000000..2927dd387 --- /dev/null +++ b/dev/trainer_rank_cost_calibration_certificate_dense-gdn-h5120-h200-bf16.json @@ -0,0 +1,83 @@ +{ + "admitted": {"device_classes":[{"capability":[9,0],"memory_class":"hbm-141g"}],"geometries":[{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4}],"param_dtypes":["torch.bfloat16"],"shapes":[[1,1,1,1],[1,2,1,1],[1,4,1,1],[2,1,1,1]]}, + "cells": [ + {"candidates":[{"cp_plan":{"max_rank_tokens":12659,"summary_ms":0.0004936009645462036,"wave_count":0},"fallback_work":882325888,"features":{"max_depth":2,"packed_tokens":12659,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":5303.195556640625,"n":8,"spread_pct":1.4776339203280155},{"cp_plan":{"max_rank_tokens":24836,"summary_ms":0.00037997961044311523,"wave_count":0},"fallback_work":1681916416,"features":{"max_depth":2,"packed_tokens":24836,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12268","median_ms":10437.79248046875,"n":8,"spread_pct":0.2942376380826529},{"cp_plan":{"max_rank_tokens":37143,"summary_ms":0.0003594905138015747,"wave_count":0},"fallback_work":2490043264,"features":{"max_depth":2,"packed_tokens":37143,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12313","median_ms":15435.33935546875,"n":8,"spread_pct":0.17101963442188503},{"cp_plan":{"max_rank_tokens":12609,"summary_ms":0.00034831464290618896,"wave_count":0},"fallback_work":896049280,"features":{"max_depth":3,"packed_tokens":12609,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":5429.99267578125,"n":8,"spread_pct":0.6007396167031979},{"cp_plan":{"max_rank_tokens":49460,"summary_ms":0.0013764947652816772,"wave_count":0},"fallback_work":3248265728,"features":{"max_depth":1,"packed_tokens":49460,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":19733.541015625,"n":8,"spread_pct":0.20979706881911975},{"cp_plan":{"max_rank_tokens":12569,"summary_ms":0.000692903995513916,"wave_count":0},"fallback_work":893422720,"features":{"max_depth":3,"packed_tokens":12569,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":5414.51806640625,"n":8,"spread_pct":0.47836883855281664},{"cp_plan":{"max_rank_tokens":12564,"summary_ms":0.0004172325134277344,"wave_count":0},"fallback_work":910100992,"features":{"max_depth":4,"packed_tokens":12564,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":5530.77099609375,"n":8,"spread_pct":0.2973068200909701}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"f8b37d8f36ed38a4d2ea74f20bfe0d15bedbdb66ffaae41eb728a1756d67d5ff","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":0,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12382,"summary_ms":0.0003948807716369629,"wave_count":0},"fallback_work":863874816,"features":{"max_depth":2,"packed_tokens":12382,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":5166.57763671875,"n":8,"spread_pct":0.5116929150713376},{"cp_plan":{"max_rank_tokens":24645,"summary_ms":0.0009406358003616333,"wave_count":0},"fallback_work":1618551424,"features":{"max_depth":1,"packed_tokens":24645,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":9934.1591796875,"n":8,"spread_pct":0.5544120129221901}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"515df421b93fe5f98e250ae3d849975dccadfd7dc00046242b5875ddf392c74f","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":1,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12546,"summary_ms":0.0007413327693939209,"wave_count":0},"fallback_work":823951616,"features":{"max_depth":1,"packed_tokens":12546,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5142.018310546875,"n":8,"spread_pct":0.4447122923132488}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"8aa85999c5d3c576d127bd2e3a44436cbd0e9062c7c92b16bcbaab0814703094","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":2,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":47253,"summary_ms":0.00042282044887542725,"wave_count":0},"fallback_work":3154037376,"features":{"max_depth":2,"packed_tokens":47253,"segment_count":6,"segments_below":[0,0,0,1,1,1,1,3]},"label":"depth_one","median_ms":18846.056640625,"n":8,"spread_pct":0.3356761573857926},{"cp_plan":{"max_rank_tokens":23150,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":1571567360,"features":{"max_depth":2,"packed_tokens":23150,"segment_count":7,"segments_below":[0,0,0,1,3,5,5,6]},"label":"minimum_effective_span_1366","median_ms":9351.30078125,"n":8,"spread_pct":0.2570038509849691},{"cp_plan":{"max_rank_tokens":21785,"summary_ms":0.00022351741790771484,"wave_count":0},"fallback_work":1498811520,"features":{"max_depth":3,"packed_tokens":21785,"segment_count":7,"segments_below":[0,0,0,1,4,5,5,6]},"label":"minimum_effective_span_305","median_ms":8919.32470703125,"n":8,"spread_pct":0.31630100920375825},{"cp_plan":{"max_rank_tokens":28799,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":1942372224,"features":{"max_depth":2,"packed_tokens":28799,"segment_count":6,"segments_below":[0,0,0,1,3,3,3,4]},"label":"minimum_effective_span_7015","median_ms":11600.98046875,"n":8,"spread_pct":0.2094381165923812},{"cp_plan":{"max_rank_tokens":37178,"summary_ms":0.0001993030309677124,"wave_count":0},"fallback_work":2492472576,"features":{"max_depth":2,"packed_tokens":37178,"segment_count":6,"segments_below":[0,0,0,1,2,2,2,3]},"label":"minimum_effective_span_8380","median_ms":14960.962890625,"n":8,"spread_pct":0.36052810876430624},{"cp_plan":{"max_rank_tokens":48469,"summary_ms":0.0010766088962554932,"wave_count":0},"fallback_work":3183323776,"features":{"max_depth":1,"packed_tokens":48469,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":18895.2841796875,"n":8,"spread_pct":0.45561550136698603},{"cp_plan":{"max_rank_tokens":22846,"summary_ms":0.00027567148208618164,"wave_count":0},"fallback_work":1568612096,"features":{"max_depth":3,"packed_tokens":22846,"segment_count":8,"segments_below":[0,0,0,2,4,6,6,7]},"label":"uniform_depth_2","median_ms":9334.0869140625,"n":8,"spread_pct":0.12621748767145788},{"cp_plan":{"max_rank_tokens":21481,"summary_ms":0.00024400651454925537,"wave_count":0},"fallback_work":1495856256,"features":{"max_depth":4,"packed_tokens":21481,"segment_count":8,"segments_below":[0,0,0,2,5,6,6,7]},"label":"uniform_depth_3","median_ms":8984.55810546875,"n":8,"spread_pct":0.4667404876259315}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"62112dd684879b09adbef34aa98e5ea3966b7de9b9e8067de5920b1d73f07197","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":3,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":30477,"summary_ms":0.00033527612686157227,"wave_count":0},"fallback_work":2052195968,"features":{"max_depth":2,"packed_tokens":30477,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,2]},"label":"depth_one","median_ms":12298.390625,"n":8,"spread_pct":0.18315722204505924},{"cp_plan":{"max_rank_tokens":19821,"summary_ms":0.00022351741790771484,"wave_count":0},"fallback_work":1352480384,"features":{"max_depth":2,"packed_tokens":19821,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_495","median_ms":8058.333984375,"n":8,"spread_pct":0.21219176478543783},{"cp_plan":{"max_rank_tokens":31465,"summary_ms":0.0007171183824539185,"wave_count":0},"fallback_work":2066510976,"features":{"max_depth":1,"packed_tokens":31465,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":12388.9052734375,"n":8,"spread_pct":0.23972431861010976},{"cp_plan":{"max_rank_tokens":19327,"summary_ms":0.0003520399332046509,"wave_count":0},"fallback_work":1337048960,"features":{"max_depth":3,"packed_tokens":19327,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,4]},"label":"uniform_depth_2","median_ms":7915.997314453125,"n":8,"spread_pct":0.34235830486284435}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"28b4492a5c1ce50958108e27b72b7afdf09941e4670076dfa879b517d9cc7575","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":4,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13557,"summary_ms":0.0004414469003677368,"wave_count":0},"fallback_work":941423232,"features":{"max_depth":2,"packed_tokens":13557,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":5732.08154296875,"n":8,"spread_pct":0.25383960189118554},{"cp_plan":{"max_rank_tokens":25902,"summary_ms":0.00029616057872772217,"wave_count":0},"fallback_work":1752045312,"features":{"max_depth":2,"packed_tokens":25902,"segment_count":6,"segments_below":[0,1,2,4,4,4,4,4]},"label":"minimum_effective_span_12406","median_ms":10851.71630859375,"n":8,"spread_pct":0.35590748241284376},{"cp_plan":{"max_rank_tokens":37995,"summary_ms":0.0002980232238769531,"wave_count":0},"fallback_work":2546120064,"features":{"max_depth":2,"packed_tokens":37995,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12426","median_ms":15878.3046875,"n":8,"spread_pct":0.2959214199012706},{"cp_plan":{"max_rank_tokens":50582,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":3372632832,"features":{"max_depth":2,"packed_tokens":50582,"segment_count":6,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12592","median_ms":21074.314453125,"n":8,"spread_pct":0.2761989310706559},{"cp_plan":{"max_rank_tokens":13367,"summary_ms":0.00029616057872772217,"wave_count":0},"fallback_work":945953664,"features":{"max_depth":3,"packed_tokens":13367,"segment_count":7,"segments_below":[1,3,5,6,6,6,6,6]},"label":"minimum_effective_span_187","median_ms":5784.398193359375,"n":8,"spread_pct":0.2823968957782491},{"cp_plan":{"max_rank_tokens":13185,"summary_ms":0.00027008354663848877,"wave_count":0},"fallback_work":934002816,"features":{"max_depth":3,"packed_tokens":13185,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_21","median_ms":5733.703369140625,"n":8,"spread_pct":0.2231783874209722},{"cp_plan":{"max_rank_tokens":63177,"summary_ms":0.0011324882507324219,"wave_count":0},"fallback_work":4149109888,"features":{"max_depth":1,"packed_tokens":63177,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":25545.7080078125,"n":8,"spread_pct":0.19297517604493017},{"cp_plan":{"max_rank_tokens":13497,"summary_ms":0.0003073364496231079,"wave_count":0},"fallback_work":954489984,"features":{"max_depth":3,"packed_tokens":13497,"segment_count":7,"segments_below":[1,2,4,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":5840.627197265625,"n":8,"spread_pct":0.16835534062340884},{"cp_plan":{"max_rank_tokens":13165,"summary_ms":0.0002942979335784912,"wave_count":0},"fallback_work":949696128,"features":{"max_depth":4,"packed_tokens":13165,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":5843.243896484375,"n":8,"spread_pct":0.22384124427219987},{"cp_plan":{"max_rank_tokens":13161,"summary_ms":0.000303611159324646,"wave_count":0},"fallback_work":966440064,"features":{"max_depth":5,"packed_tokens":13161,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":5970.395751953125,"n":8,"spread_pct":0.24316757229201486}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"7241823ac1b67d127c5921d771dda170dcc9d4886feffaff3c2520113cd01a08","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":5,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.9118811041116714,"wave_count":2},"fallback_work":881914944,"features":{"max_depth":2,"packed_tokens":12659,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":4748.375732421875,"n":8,"spread_pct":0.22390468944792852},{"cp_plan":{"max_rank_tokens":15872,"summary_ms":12.262728065252304,"wave_count":1},"fallback_work":1680920896,"features":{"max_depth":2,"packed_tokens":24836,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12268","median_ms":5738.371826171875,"n":8,"spread_pct":0.3630218396399546},{"cp_plan":{"max_rank_tokens":21504,"summary_ms":7.400814443826675,"wave_count":1},"fallback_work":2488457088,"features":{"max_depth":2,"packed_tokens":37143,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12313","median_ms":9287.7158203125,"n":8,"spread_pct":0.5823384193313981},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.97638264298439,"wave_count":2},"fallback_work":895706272,"features":{"max_depth":3,"packed_tokens":12609,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":4740.755126953125,"n":8,"spread_pct":0.37062265460845284},{"cp_plan":{"max_rank_tokens":26420,"summary_ms":7.924016565084457,"wave_count":1},"fallback_work":3246022720,"features":{"max_depth":1,"packed_tokens":49460,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":10589.4736328125,"n":8,"spread_pct":0.1772562873600986},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.803320556879044,"wave_count":2},"fallback_work":893081632,"features":{"max_depth":3,"packed_tokens":12569,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":4746.56689453125,"n":8,"spread_pct":0.15272140051058697},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.8084409683942795,"wave_count":2},"fallback_work":909825600,"features":{"max_depth":4,"packed_tokens":12564,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":4744.302978515625,"n":8,"spread_pct":0.11390100122654304}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"f8b37d8f36ed38a4d2ea74f20bfe0d15bedbdb66ffaae41eb728a1756d67d5ff","shape":[1,2,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":0,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.701810121536255,"wave_count":2},"fallback_work":863411552,"features":{"max_depth":2,"packed_tokens":12382,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":4735.50927734375,"n":8,"spread_pct":0.6272218403119798},{"cp_plan":{"max_rank_tokens":15872,"summary_ms":12.807516381144524,"wave_count":1},"fallback_work":1617434080,"features":{"max_depth":1,"packed_tokens":24645,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5636.36376953125,"n":8,"spread_pct":0.7719127841817481}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"515df421b93fe5f98e250ae3d849975dccadfd7dc00046242b5875ddf392c74f","shape":[1,2,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":1,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.5871643126010895,"wave_count":2},"fallback_work":823382176,"features":{"max_depth":1,"packed_tokens":12546,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":4775.349609375,"n":8,"spread_pct":0.4687156822205835}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"8aa85999c5d3c576d127bd2e3a44436cbd0e9062c7c92b16bcbaab0814703094","shape":[1,2,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":2,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":24213,"summary_ms":9.730514138936996,"wave_count":1},"fallback_work":3151998688,"features":{"max_depth":2,"packed_tokens":47253,"segment_count":6,"segments_below":[0,0,0,1,1,1,1,3]},"label":"depth_one","median_ms":10813.5771484375,"n":8,"spread_pct":0.19974527465336503},{"cp_plan":{"max_rank_tokens":11776,"summary_ms":7.137704640626907,"wave_count":1},"fallback_work":1570751072,"features":{"max_depth":2,"packed_tokens":23150,"segment_count":7,"segments_below":[0,0,0,1,3,5,5,6]},"label":"minimum_effective_span_1366","median_ms":5107.97412109375,"n":8,"spread_pct":0.5080999368041562},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":5.670497193932533,"wave_count":1},"fallback_work":1498093600,"features":{"max_depth":3,"packed_tokens":21785,"segment_count":7,"segments_below":[0,0,0,1,4,5,5,6]},"label":"minimum_effective_span_305","median_ms":4860.65869140625,"n":8,"spread_pct":0.5506182098800468},{"cp_plan":{"max_rank_tokens":15360,"summary_ms":12.829218059778214,"wave_count":1},"fallback_work":1941252096,"features":{"max_depth":2,"packed_tokens":28799,"segment_count":6,"segments_below":[0,0,0,1,3,3,3,4]},"label":"minimum_effective_span_7015","median_ms":6442.249755859375,"n":8,"spread_pct":0.8772805107578583},{"cp_plan":{"max_rank_tokens":18746,"summary_ms":7.001526653766632,"wave_count":1},"fallback_work":2490917408,"features":{"max_depth":2,"packed_tokens":37178,"segment_count":6,"segments_below":[0,0,0,1,2,2,2,3]},"label":"minimum_effective_span_8380","median_ms":8255.71484375,"n":8,"spread_pct":0.6920632156796689},{"cp_plan":{"max_rank_tokens":25941,"summary_ms":7.394980639219284,"wave_count":1},"fallback_work":3181161184,"features":{"max_depth":1,"packed_tokens":48469,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":11122.4677734375,"n":8,"spread_pct":0.33984210064667986},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":6.164057180285454,"wave_count":1},"fallback_work":1567875936,"features":{"max_depth":3,"packed_tokens":22846,"segment_count":8,"segments_below":[0,0,0,2,4,6,6,7]},"label":"uniform_depth_2","median_ms":5273.21240234375,"n":8,"spread_pct":1.130464920771155},{"cp_plan":{"max_rank_tokens":11776,"summary_ms":6.091278046369553,"wave_count":1},"fallback_work":1495218464,"features":{"max_depth":4,"packed_tokens":21481,"segment_count":8,"segments_below":[0,0,0,2,5,6,6,7]},"label":"uniform_depth_3","median_ms":5007.601318359375,"n":8,"spread_pct":0.541998285321607}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"62112dd684879b09adbef34aa98e5ea3966b7de9b9e8067de5920b1d73f07197","shape":[1,2,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":3,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":18432,"summary_ms":7.784789428114891,"wave_count":1},"fallback_work":2050896992,"features":{"max_depth":2,"packed_tokens":30477,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,2]},"label":"depth_one","median_ms":7792.415771484375,"n":8,"spread_pct":0.3562470833221718},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":6.154848262667656,"wave_count":1},"fallback_work":1351692896,"features":{"max_depth":2,"packed_tokens":19821,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_495","median_ms":4805.787841796875,"n":8,"spread_pct":0.28614381858788746},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":6.0643404722213745,"wave_count":1},"fallback_work":2065099040,"features":{"max_depth":1,"packed_tokens":31465,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":6879.4033203125,"n":8,"spread_pct":0.39708232326766846},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":6.4881350845098495,"wave_count":1},"fallback_work":1336350720,"features":{"max_depth":3,"packed_tokens":19327,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,4]},"label":"uniform_depth_2","median_ms":4738.11572265625,"n":8,"spread_pct":0.2886127204979646}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"28b4492a5c1ce50958108e27b72b7afdf09941e4670076dfa879b517d9cc7575","shape":[1,2,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":4,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10752,"summary_ms":6.026476621627808,"wave_count":2},"fallback_work":941001952,"features":{"max_depth":2,"packed_tokens":13557,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":4968.552734375,"n":8,"spread_pct":0.6115219196988559},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":7.254950702190399,"wave_count":1},"fallback_work":1751031392,"features":{"max_depth":2,"packed_tokens":25902,"segment_count":6,"segments_below":[0,1,2,4,4,4,4,4]},"label":"minimum_effective_span_12406","median_ms":5768.194091796875,"n":8,"spread_pct":0.3121830495329338},{"cp_plan":{"max_rank_tokens":22016,"summary_ms":7.527731359004974,"wave_count":1},"fallback_work":2544525760,"features":{"max_depth":2,"packed_tokens":37995,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12426","median_ms":9627.4921875,"n":8,"spread_pct":0.1876644764870135},{"cp_plan":{"max_rank_tokens":26518,"summary_ms":9.276259690523148,"wave_count":1},"fallback_work":3370434272,"features":{"max_depth":2,"packed_tokens":50582,"segment_count":6,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12592","median_ms":10856.0390625,"n":8,"spread_pct":0.184616065510772},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":6.7081451416015625,"wave_count":2},"fallback_work":945607040,"features":{"max_depth":3,"packed_tokens":13367,"segment_count":7,"segments_below":[1,3,5,6,6,6,6,6]},"label":"minimum_effective_span_187","median_ms":4881.4697265625,"n":8,"spread_pct":0.4794318437570332},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.841750651597977,"wave_count":2},"fallback_work":933664928,"features":{"max_depth":3,"packed_tokens":13185,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_21","median_ms":4880.201171875,"n":8,"spread_pct":0.644454654349146},{"cp_plan":{"max_rank_tokens":34304,"summary_ms":7.157968357205391,"wave_count":1},"fallback_work":4146241312,"features":{"max_depth":1,"packed_tokens":63177,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":14612.0390625,"n":8,"spread_pct":0.1255787046319361},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":5.938926711678505,"wave_count":2},"fallback_work":954137120,"features":{"max_depth":3,"packed_tokens":13497,"segment_count":7,"segments_below":[1,2,4,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":4980.02294921875,"n":8,"spread_pct":1.1602901318519983},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":5.8663710951805115,"wave_count":2},"fallback_work":949424736,"features":{"max_depth":4,"packed_tokens":13165,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":4882.7724609375,"n":8,"spread_pct":0.6801055768657303},{"cp_plan":{"max_rank_tokens":10240,"summary_ms":6.0594528913497925,"wave_count":2},"fallback_work":966234400,"features":{"max_depth":5,"packed_tokens":13161,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":4881.28173828125,"n":8,"spread_pct":0.7905978524267357}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"7241823ac1b67d127c5921d771dda170dcc9d4886feffaff3c2520113cd01a08","shape":[1,2,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":5,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":7680,"summary_ms":7.146457210183144,"wave_count":1},"fallback_work":671712288,"features":{"max_depth":2,"packed_tokens":9433,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":2497.5184326171875,"n":8,"spread_pct":0.4294203948831826},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.117817014455795,"wave_count":2},"fallback_work":563001760,"features":{"max_depth":6,"packed_tokens":6738,"segment_count":17,"segments_below":[6,11,15,16,16,16,16,17]},"label":"minimum_effective_span_105","median_ms":2480.2047119140625,"n":8,"spread_pct":0.80799939481747},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.150024011731148,"wave_count":2},"fallback_work":565626400,"features":{"max_depth":6,"packed_tokens":6778,"segment_count":17,"segments_below":[5,10,15,16,16,16,16,17]},"label":"minimum_effective_span_113","median_ms":2476.401611328125,"n":8,"spread_pct":1.3388290899418642},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.538204848766327,"wave_count":2},"fallback_work":572122464,"features":{"max_depth":6,"packed_tokens":6877,"segment_count":17,"segments_below":[5,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":2479.632080078125,"n":8,"spread_pct":26.26434132851701},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.509967148303986,"wave_count":2},"fallback_work":569026464,"features":{"max_depth":5,"packed_tokens":7090,"segment_count":16,"segments_below":[3,8,14,15,15,15,15,16]},"label":"minimum_effective_span_164","median_ms":2400.36669921875,"n":8,"spread_pct":0.6273150403655366},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.113691091537476,"wave_count":2},"fallback_work":571913568,"features":{"max_depth":5,"packed_tokens":7134,"segment_count":16,"segments_below":[4,7,14,15,15,15,15,16]},"label":"minimum_effective_span_195","median_ms":2403.451904296875,"n":8,"spread_pct":0.5093890125276562},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.8766920119524,"wave_count":2},"fallback_work":573443456,"features":{"max_depth":4,"packed_tokens":7415,"segment_count":16,"segments_below":[3,5,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":2321.4796142578125,"n":8,"spread_pct":0.31583440168799776},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.279893219470978,"wave_count":2},"fallback_work":587124512,"features":{"max_depth":4,"packed_tokens":7626,"segment_count":15,"segments_below":[3,6,10,14,14,14,14,15]},"label":"minimum_effective_span_245","median_ms":2329.2330322265625,"n":8,"spread_pct":0.3854709039746551},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.879942491650581,"wave_count":2},"fallback_work":592111328,"features":{"max_depth":4,"packed_tokens":7702,"segment_count":15,"segments_below":[3,5,9,14,14,14,14,15]},"label":"minimum_effective_span_281","median_ms":2340.760009765625,"n":8,"spread_pct":0.6290733605673822},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.249995738267899,"wave_count":1},"fallback_work":638203968,"features":{"max_depth":12,"packed_tokens":6323,"segment_count":23,"segments_below":[16,20,22,22,22,22,22,23]},"label":"minimum_effective_span_32","median_ms":3610.1534423828125,"n":8,"spread_pct":0.6006956630058513},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.928931921720505,"wave_count":2},"fallback_work":597446944,"features":{"max_depth":3,"packed_tokens":8041,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_340","median_ms":2382.1524658203125,"n":8,"spread_pct":0.2760600746323534},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":5.217384546995163,"wave_count":1},"fallback_work":602171296,"features":{"max_depth":3,"packed_tokens":8113,"segment_count":15,"segments_below":[1,4,8,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":2386.14892578125,"n":8,"spread_pct":0.7984209814413548},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.280191078782082,"wave_count":1},"fallback_work":623494016,"features":{"max_depth":11,"packed_tokens":6359,"segment_count":22,"segments_below":[15,19,21,21,21,21,21,22]},"label":"minimum_effective_span_37","median_ms":3446.0343017578125,"n":8,"spread_pct":1.2192315293355669},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.122261121869087,"wave_count":2},"fallback_work":594467808,"features":{"max_depth":9,"packed_tokens":6437,"segment_count":20,"segments_below":[9,17,19,19,19,19,19,20]},"label":"minimum_effective_span_40","median_ms":3040.7894287109375,"n":8,"spread_pct":1.284376202195836},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.735213100910187,"wave_count":1},"fallback_work":620412544,"features":{"max_depth":3,"packed_tokens":8391,"segment_count":15,"segments_below":[1,3,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":2412.2216796875,"n":8,"spread_pct":0.5945067334921654},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":6.166618317365646,"wave_count":1},"fallback_work":968624608,"features":{"max_depth":2,"packed_tokens":13958,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4955","median_ms":3258.091796875,"n":8,"spread_pct":0.5503275198828878},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":5.963720381259918,"wave_count":1},"fallback_work":1260812736,"features":{"max_depth":2,"packed_tokens":18411,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4994","median_ms":4775.560546875,"n":8,"spread_pct":0.5292654208214268},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":7.072145119309425,"wave_count":1},"fallback_work":1568945472,"features":{"max_depth":2,"packed_tokens":23107,"segment_count":14,"segments_below":[2,3,5,7,10,10,10,14]},"label":"minimum_effective_span_5048","median_ms":4940.603759765625,"n":8,"spread_pct":0.534484034042485},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":7.5146593153476715,"wave_count":1},"fallback_work":1886395680,"features":{"max_depth":2,"packed_tokens":27945,"segment_count":14,"segments_below":[2,3,5,6,9,9,9,14]},"label":"minimum_effective_span_5087","median_ms":6378.607177734375,"n":8,"spread_pct":0.35678341471536384},{"cp_plan":{"max_rank_tokens":19456,"summary_ms":6.539067253470421,"wave_count":1},"fallback_work":2207914080,"features":{"max_depth":2,"packed_tokens":32845,"segment_count":14,"segments_below":[1,3,4,6,8,8,8,14]},"label":"minimum_effective_span_5118","median_ms":6997.284423828125,"n":8,"spread_pct":0.2439562473960584},{"cp_plan":{"max_rank_tokens":22016,"summary_ms":8.767664432525635,"wave_count":1},"fallback_work":2526020448,"features":{"max_depth":2,"packed_tokens":37693,"segment_count":14,"segments_below":[2,2,3,6,7,7,7,14]},"label":"minimum_effective_span_5149","median_ms":8172.88330078125,"n":8,"spread_pct":0.24995720831529808},{"cp_plan":{"max_rank_tokens":24064,"summary_ms":7.65908882021904,"wave_count":1},"fallback_work":2855281536,"features":{"max_depth":2,"packed_tokens":42711,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_5199","median_ms":8770.14599609375,"n":8,"spread_pct":0.3653307376726765},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":7.356669753789902,"wave_count":1},"fallback_work":643542304,"features":{"max_depth":3,"packed_tokens":8746,"segment_count":14,"segments_below":[2,4,7,10,13,13,13,14]},"label":"minimum_effective_span_522","median_ms":2381.550048828125,"n":8,"spread_pct":24.134180107103877},{"cp_plan":{"max_rank_tokens":27136,"summary_ms":7.1244314312934875,"wave_count":2},"fallback_work":3180868128,"features":{"max_depth":2,"packed_tokens":47673,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5235","median_ms":10003.04638671875,"n":8,"spread_pct":0.4633842258698632},{"cp_plan":{"max_rank_tokens":29184,"summary_ms":8.530354127287865,"wave_count":1},"fallback_work":3510654144,"features":{"max_depth":2,"packed_tokens":52699,"segment_count":14,"segments_below":[0,1,2,4,4,4,4,14]},"label":"minimum_effective_span_5303","median_ms":10574.4208984375,"n":8,"spread_pct":0.2373338448369153},{"cp_plan":{"max_rank_tokens":32256,"summary_ms":7.926104590296745,"wave_count":2},"fallback_work":3853957056,"features":{"max_depth":2,"packed_tokens":57931,"segment_count":14,"segments_below":[0,0,3,3,3,3,3,14]},"label":"minimum_effective_span_5395","median_ms":11830.2685546875,"n":8,"spread_pct":0.394859088862281},{"cp_plan":{"max_rank_tokens":34304,"summary_ms":9.516201913356781,"wave_count":1},"fallback_work":4202148480,"features":{"max_depth":2,"packed_tokens":63240,"segment_count":13,"segments_below":[1,1,1,1,1,1,1,13]},"label":"minimum_effective_span_5476","median_ms":12395.46142578125,"n":8,"spread_pct":0.3230138929457114},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.112100392580032,"wave_count":2},"fallback_work":585335136,"features":{"max_depth":8,"packed_tokens":6558,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_71","median_ms":2857.6328125,"n":8,"spread_pct":2.6662078260798245},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.175100639462471,"wave_count":2},"fallback_work":573577984,"features":{"max_depth":7,"packed_tokens":6639,"segment_count":18,"segments_below":[6,14,16,17,17,17,17,18]},"label":"minimum_effective_span_82","median_ms":2669.924560546875,"n":8,"spread_pct":1.7839137140589927},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.199587136507034,"wave_count":2},"fallback_work":574759072,"features":{"max_depth":7,"packed_tokens":6657,"segment_count":18,"segments_below":[6,14,16,17,17,17,17,18]},"label":"minimum_effective_span_87","median_ms":2667.326171875,"n":8,"spread_pct":1.621748576340861},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.277529522776604,"wave_count":2},"fallback_work":575480768,"features":{"max_depth":7,"packed_tokens":6668,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_93","median_ms":2662.127685546875,"n":8,"spread_pct":0.66846568828626},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.2470434457063675,"wave_count":2},"fallback_work":573971680,"features":{"max_depth":7,"packed_tokens":6645,"segment_count":18,"segments_below":[7,12,17,17,17,17,17,18]},"label":"minimum_effective_span_94","median_ms":2678.144287109375,"n":8,"spread_pct":1.1191211702730772},{"cp_plan":{"max_rank_tokens":37376,"summary_ms":6.816597655415535,"wave_count":1},"fallback_work":4521825696,"features":{"max_depth":1,"packed_tokens":68881,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":13772.62841796875,"n":8,"spread_pct":0.26191981002287323},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.416098982095718,"wave_count":2},"fallback_work":640783552,"features":{"max_depth":11,"packed_tokens":6620,"segment_count":23,"segments_below":[16,19,20,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":3439.69140625,"n":8,"spread_pct":1.2284924042740935},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.412464961409569,"wave_count":2},"fallback_work":647225888,"features":{"max_depth":12,"packed_tokens":6458,"segment_count":24,"segments_below":[16,20,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":3640.7662353515625,"n":8,"spread_pct":0.8989726357517581},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.248043850064278,"wave_count":1},"fallback_work":653241920,"features":{"max_depth":13,"packed_tokens":6292,"segment_count":24,"segments_below":[17,21,23,23,23,23,23,24]},"label":"uniform_depth_12","median_ms":3798.855224609375,"n":8,"spread_pct":1.045995108903786},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":7.158270105719566,"wave_count":1},"fallback_work":660635072,"features":{"max_depth":3,"packed_tokens":9004,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":2544.0023193359375,"n":8,"spread_pct":0.5832978694767252},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.5311044454574585,"wave_count":1},"fallback_work":642274560,"features":{"max_depth":4,"packed_tokens":8464,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":2491.577392578125,"n":8,"spread_pct":25.40958057816074},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":5.867870524525642,"wave_count":2},"fallback_work":636315552,"features":{"max_depth":5,"packed_tokens":8113,"segment_count":17,"segments_below":[8,9,11,13,16,16,16,17]},"label":"uniform_depth_4","median_ms":2580.2855224609375,"n":8,"spread_pct":1.244619861414433},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":7.038539275527,"wave_count":1},"fallback_work":637114912,"features":{"max_depth":6,"packed_tokens":7865,"segment_count":18,"segments_below":[10,11,13,14,17,17,17,18]},"label":"uniform_depth_5","median_ms":2699.95263671875,"n":8,"spread_pct":0.8940134402675973},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.0463473200798035,"wave_count":2},"fallback_work":639948288,"features":{"max_depth":7,"packed_tokens":7648,"segment_count":19,"segments_below":[11,13,14,16,18,18,18,19]},"label":"uniform_depth_6","median_ms":2804.6590576171875,"n":8,"spread_pct":1.0963290148267957},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.905358120799065,"wave_count":2},"fallback_work":637335616,"features":{"max_depth":8,"packed_tokens":7348,"segment_count":20,"segments_below":[14,14,15,18,19,19,19,20]},"label":"uniform_depth_7","median_ms":2980.7650146484375,"n":8,"spread_pct":0.5391583986961977},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.044642835855484,"wave_count":2},"fallback_work":642596864,"features":{"max_depth":9,"packed_tokens":7168,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":3153.40234375,"n":8,"spread_pct":1.2537843549440026},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.769990548491478,"wave_count":2},"fallback_work":641821440,"features":{"max_depth":10,"packed_tokens":6896,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":3257.9508056640625,"n":8,"spread_pct":2.1493616896434964}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"22fe33a263cacb6bc86ab088fe07af0a0fb8e68bfdc85c518a57ec37e403d138","shape":[1,2,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":6,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.139691919088364,"wave_count":2},"fallback_work":882004320,"features":{"max_depth":2,"packed_tokens":12659,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1863.1535034179688,"n":8,"spread_pct":1.1450994723924173},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":8.911814540624619,"wave_count":3},"fallback_work":1680718048,"features":{"max_depth":2,"packed_tokens":24836,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12268","median_ms":4609.49365234375,"n":8,"spread_pct":0.31274618672421883},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":10.849203914403915,"wave_count":2},"fallback_work":2487958848,"features":{"max_depth":2,"packed_tokens":37143,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12313","median_ms":5083.161376953125,"n":8,"spread_pct":0.2490222557637428},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.876382976770401,"wave_count":2},"fallback_work":895928032,"features":{"max_depth":3,"packed_tokens":12609,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1940.5178833007812,"n":8,"spread_pct":0.4509483008211284},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":22.069288417696953,"wave_count":1},"fallback_work":3245097824,"features":{"max_depth":1,"packed_tokens":49460,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5430.927001953125,"n":8,"spread_pct":0.2482886438199337},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.9233570247888565,"wave_count":2},"fallback_work":893304352,"features":{"max_depth":3,"packed_tokens":12569,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1949.6373901367188,"n":8,"spread_pct":0.8652763360975297},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.982531398534775,"wave_count":2},"fallback_work":910179424,"features":{"max_depth":4,"packed_tokens":12564,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":2078.4095458984375,"n":8,"spread_pct":1.1624700588752797}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"f8b37d8f36ed38a4d2ea74f20bfe0d15bedbdb66ffaae41eb728a1756d67d5ff","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":0,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":5.095181986689568,"wave_count":2},"fallback_work":863376640,"features":{"max_depth":2,"packed_tokens":12382,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1840.7072143554688,"n":8,"spread_pct":1.1601309589743694},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":8.774839341640472,"wave_count":3},"fallback_work":1616973760,"features":{"max_depth":1,"packed_tokens":24645,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":4570.117919921875,"n":8,"spread_pct":0.792288004996794}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"515df421b93fe5f98e250ae3d849975dccadfd7dc00046242b5875ddf392c74f","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":1,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.9581266939640045,"wave_count":2},"fallback_work":823146720,"features":{"max_depth":1,"packed_tokens":12546,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1776.1741333007812,"n":8,"spread_pct":0.4891543153896226}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"8aa85999c5d3c576d127bd2e3a44436cbd0e9062c7c92b16bcbaab0814703094","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":2,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":14336,"summary_ms":5.45392744243145,"wave_count":1},"fallback_work":3151323456,"features":{"max_depth":2,"packed_tokens":47253,"segment_count":6,"segments_below":[0,0,0,1,1,1,1,3]},"label":"depth_one","median_ms":6330.8154296875,"n":8,"spread_pct":0.3697964831310145},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":5.357343703508377,"wave_count":2},"fallback_work":1570785408,"features":{"max_depth":2,"packed_tokens":23150,"segment_count":7,"segments_below":[0,0,0,1,3,5,5,6]},"label":"minimum_effective_span_1366","median_ms":4235.478515625,"n":8,"spread_pct":0.3164650356341636},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":6.099650636315346,"wave_count":2},"fallback_work":1498226208,"features":{"max_depth":3,"packed_tokens":21785,"segment_count":7,"segments_below":[0,0,0,1,4,5,5,6]},"label":"minimum_effective_span_305","median_ms":4169.562744140625,"n":8,"spread_pct":0.1949231107667411},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":4.570197314023972,"wave_count":3},"fallback_work":1941085184,"features":{"max_depth":2,"packed_tokens":28799,"segment_count":6,"segments_below":[0,0,0,1,3,3,3,4]},"label":"minimum_effective_span_7015","median_ms":4325.090087890625,"n":8,"spread_pct":0.3421284408081953},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":4.305066540837288,"wave_count":1},"fallback_work":2490484000,"features":{"max_depth":2,"packed_tokens":37178,"segment_count":6,"segments_below":[0,0,0,1,2,2,2,3]},"label":"minimum_effective_span_8380","median_ms":4421.6396484375,"n":8,"spread_pct":0.0943844405146122},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":25.78931301832199,"wave_count":1},"fallback_work":3180325696,"features":{"max_depth":1,"packed_tokens":48469,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":6490.125244140625,"n":8,"spread_pct":0.3470190760429954},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":6.529252976179123,"wave_count":1},"fallback_work":1568048640,"features":{"max_depth":3,"packed_tokens":22846,"segment_count":8,"segments_below":[0,0,0,2,4,6,6,7]},"label":"uniform_depth_2","median_ms":4292.877197265625,"n":8,"spread_pct":0.3998492915994284},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.203071817755699,"wave_count":2},"fallback_work":1495489440,"features":{"max_depth":4,"packed_tokens":21481,"segment_count":8,"segments_below":[0,0,0,2,5,6,6,7]},"label":"uniform_depth_3","median_ms":4214.74853515625,"n":8,"spread_pct":0.5106927230167966}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"62112dd684879b09adbef34aa98e5ea3966b7de9b9e8067de5920b1d73f07197","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":3,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12288,"summary_ms":12.265888974070549,"wave_count":1},"fallback_work":2050493312,"features":{"max_depth":2,"packed_tokens":30477,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,2]},"label":"depth_one","median_ms":4584.26708984375,"n":8,"spread_pct":0.43853281737310834},{"cp_plan":{"max_rank_tokens":8192,"summary_ms":7.276095449924469,"wave_count":3},"fallback_work":1351544960,"features":{"max_depth":2,"packed_tokens":19821,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_495","median_ms":4065.8160400390625,"n":8,"spread_pct":0.3186161295925366},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":16.148291528224945,"wave_count":2},"fallback_work":2064540576,"features":{"max_depth":1,"packed_tokens":31465,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":4608.88525390625,"n":8,"spread_pct":0.1669668927747302},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":7.002759724855423,"wave_count":2},"fallback_work":1336345600,"features":{"max_depth":3,"packed_tokens":19327,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,4]},"label":"uniform_depth_2","median_ms":4036.903076171875,"n":8,"spread_pct":0.6220625086369013}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"28b4492a5c1ce50958108e27b72b7afdf09941e4670076dfa879b517d9cc7575","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":4,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.353856831789017,"wave_count":2},"fallback_work":941135424,"features":{"max_depth":2,"packed_tokens":13557,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":2017.5435180664062,"n":8,"spread_pct":0.6528004034919161},{"cp_plan":{"max_rank_tokens":9216,"summary_ms":6.472637876868248,"wave_count":2},"fallback_work":1750868608,"features":{"max_depth":2,"packed_tokens":25902,"segment_count":6,"segments_below":[0,1,2,4,4,4,4,4]},"label":"minimum_effective_span_12406","median_ms":4747.364501953125,"n":8,"spread_pct":0.4033179387926226},{"cp_plan":{"max_rank_tokens":11264,"summary_ms":10.94856671988964,"wave_count":2},"fallback_work":2544072608,"features":{"max_depth":2,"packed_tokens":37995,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12426","median_ms":5196.100341796875,"n":8,"spread_pct":0.2740091998266992},{"cp_plan":{"max_rank_tokens":15872,"summary_ms":16.52071252465248,"wave_count":1},"fallback_work":3369679168,"features":{"max_depth":2,"packed_tokens":50582,"segment_count":6,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12592","median_ms":5860.517333984375,"n":8,"spread_pct":0.22847212412213466},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.309911444783211,"wave_count":3},"fallback_work":945876032,"features":{"max_depth":3,"packed_tokens":13367,"segment_count":7,"segments_below":[1,3,5,6,6,6,6,6]},"label":"minimum_effective_span_187","median_ms":2049.7174072265625,"n":8,"spread_pct":1.64935560321917},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.095819175243378,"wave_count":2},"fallback_work":933938400,"features":{"max_depth":3,"packed_tokens":13185,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_21","median_ms":2073.2747802734375,"n":8,"spread_pct":0.49457536201002966},{"cp_plan":{"max_rank_tokens":18944,"summary_ms":18.996303901076317,"wave_count":3},"fallback_work":4145052832,"features":{"max_depth":1,"packed_tokens":63177,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":7177.760009765625,"n":8,"spread_pct":0.282699670630428},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":5.374878644943237,"wave_count":2},"fallback_work":954403104,"features":{"max_depth":3,"packed_tokens":13497,"segment_count":7,"segments_below":[1,2,4,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":4296.199951171875,"n":8,"spread_pct":0.38919750744932374},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.1304643750190735,"wave_count":2},"fallback_work":949829760,"features":{"max_depth":4,"packed_tokens":13165,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":4296.63525390625,"n":8,"spread_pct":0.41714827497761525},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":6.2716044485569,"wave_count":2},"fallback_work":966770592,"features":{"max_depth":5,"packed_tokens":13161,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":4302.14111328125,"n":8,"spread_pct":0.4171251483267382}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"7241823ac1b67d127c5921d771dda170dcc9d4886feffaff3c2520113cd01a08","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":5,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":4096,"summary_ms":5.390245467424393,"wave_count":1},"fallback_work":672469024,"features":{"max_depth":2,"packed_tokens":9433,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1948.229248046875,"n":8,"spread_pct":0.4063686137263819},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.5858433693647385,"wave_count":2},"fallback_work":564281952,"features":{"max_depth":6,"packed_tokens":6738,"segment_count":17,"segments_below":[6,11,15,16,16,16,16,17]},"label":"minimum_effective_span_105","median_ms":1997.0381469726562,"n":8,"spread_pct":3.0769933244340457},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.656577318906784,"wave_count":2},"fallback_work":566905632,"features":{"max_depth":6,"packed_tokens":6778,"segment_count":17,"segments_below":[5,10,15,16,16,16,16,17]},"label":"minimum_effective_span_113","median_ms":1989.436279296875,"n":8,"spread_pct":2.185732968728505},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.60272079706192,"wave_count":2},"fallback_work":573399296,"features":{"max_depth":6,"packed_tokens":6877,"segment_count":17,"segments_below":[5,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":2044.5015869140625,"n":8,"spread_pct":2.0101910869097557},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.5241935402154922,"wave_count":2},"fallback_work":570167136,"features":{"max_depth":5,"packed_tokens":7090,"segment_count":16,"segments_below":[3,8,14,15,15,15,15,16]},"label":"minimum_effective_span_164","median_ms":1918.7649536132812,"n":8,"spread_pct":1.8102273925555028},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.525247797369957,"wave_count":2},"fallback_work":573053184,"features":{"max_depth":5,"packed_tokens":7134,"segment_count":16,"segments_below":[4,7,14,15,15,15,15,16]},"label":"minimum_effective_span_195","median_ms":1927.0569458007812,"n":8,"spread_pct":2.264473574981279},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.5460684448480606,"wave_count":2},"fallback_work":574510656,"features":{"max_depth":4,"packed_tokens":7415,"segment_count":16,"segments_below":[3,5,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":1887.3501586914062,"n":8,"spread_pct":34.716455042462016},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.5750046372413635,"wave_count":1},"fallback_work":588121248,"features":{"max_depth":4,"packed_tokens":7626,"segment_count":15,"segments_below":[3,6,10,14,14,14,14,15]},"label":"minimum_effective_span_245","median_ms":1881.9150390625,"n":8,"spread_pct":0.8360377690403258},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.5688281059265137,"wave_count":1},"fallback_work":593106240,"features":{"max_depth":4,"packed_tokens":7702,"segment_count":15,"segments_below":[3,5,9,14,14,14,14,15]},"label":"minimum_effective_span_281","median_ms":1882.1478271484375,"n":8,"spread_pct":1.1324796649232045},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.856224939227104,"wave_count":2},"fallback_work":640280416,"features":{"max_depth":12,"packed_tokens":6323,"segment_count":23,"segments_below":[16,20,22,22,22,22,22,23]},"label":"minimum_effective_span_32","median_ms":3156.1085205078125,"n":8,"spread_pct":2.785072242370394},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":6.043843924999237,"wave_count":2},"fallback_work":598368160,"features":{"max_depth":3,"packed_tokens":8041,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_340","median_ms":1882.0946655273438,"n":8,"spread_pct":1.0551092174373147},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.520231693983078,"wave_count":2},"fallback_work":603090784,"features":{"max_depth":3,"packed_tokens":8113,"segment_count":15,"segments_below":[1,4,8,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":1889.3282470703125,"n":8,"spread_pct":0.885505341118201},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.805840387940407,"wave_count":2},"fallback_work":625438528,"features":{"max_depth":11,"packed_tokens":6359,"segment_count":22,"segments_below":[15,19,21,21,21,21,21,22]},"label":"minimum_effective_span_37","median_ms":2986.804443359375,"n":8,"spread_pct":2.2800810629370676},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.7631914019584656,"wave_count":2},"fallback_work":596148416,"features":{"max_depth":9,"packed_tokens":6437,"segment_count":20,"segments_below":[9,17,19,19,19,19,19,20]},"label":"minimum_effective_span_40","median_ms":2588.97509765625,"n":8,"spread_pct":3.7394706859634854},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":3.618253394961357,"wave_count":1},"fallback_work":621325248,"features":{"max_depth":3,"packed_tokens":8391,"segment_count":15,"segments_below":[1,3,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":1941.9896240234375,"n":8,"spread_pct":0.7371590239577783},{"cp_plan":{"max_rank_tokens":5120,"summary_ms":3.902297466993332,"wave_count":2},"fallback_work":969272768,"features":{"max_depth":2,"packed_tokens":13958,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4955","median_ms":2089.268798828125,"n":8,"spread_pct":0.8806862168271762},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":4.083693027496338,"wave_count":1},"fallback_work":1261353888,"features":{"max_depth":2,"packed_tokens":18411,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4994","median_ms":2223.3804931640625,"n":8,"spread_pct":0.48260856125911955},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":4.387432709336281,"wave_count":1},"fallback_work":1569373920,"features":{"max_depth":2,"packed_tokens":23107,"segment_count":14,"segments_below":[2,3,5,7,10,10,10,14]},"label":"minimum_effective_span_5048","median_ms":2564.2584228515625,"n":8,"spread_pct":1.1697955932964255},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":4.775026813149452,"wave_count":3},"fallback_work":1886708128,"features":{"max_depth":2,"packed_tokens":27945,"segment_count":14,"segments_below":[2,3,5,6,9,9,9,14]},"label":"minimum_effective_span_5087","median_ms":3762.8914794921875,"n":8,"spread_pct":0.27064511657467116},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":5.218587815761566,"wave_count":1},"fallback_work":2208108928,"features":{"max_depth":2,"packed_tokens":32845,"segment_count":14,"segments_below":[1,3,4,6,8,8,8,14]},"label":"minimum_effective_span_5118","median_ms":4443.052734375,"n":8,"spread_pct":0.6164491351710866},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":5.597824230790138,"wave_count":1},"fallback_work":2526098944,"features":{"max_depth":2,"packed_tokens":37693,"segment_count":14,"segments_below":[2,2,3,6,7,7,7,14]},"label":"minimum_effective_span_5149","median_ms":4487.47802734375,"n":8,"spread_pct":0.8883769995827148},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":6.011003628373146,"wave_count":1},"fallback_work":2855239488,"features":{"max_depth":2,"packed_tokens":42711,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_5199","median_ms":4845.587646484375,"n":8,"spread_pct":0.6146256911503383},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":3.6177169531583786,"wave_count":1},"fallback_work":644381088,"features":{"max_depth":3,"packed_tokens":8746,"segment_count":14,"segments_below":[2,4,7,10,13,13,13,14]},"label":"minimum_effective_span_522","median_ms":1944.7896728515625,"n":8,"spread_pct":1.1140795433411226},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":10.18480584025383,"wave_count":1},"fallback_work":3180707104,"features":{"max_depth":2,"packed_tokens":47673,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5235","median_ms":6020.524658203125,"n":8,"spread_pct":0.5432831513975083},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":6.853988394141197,"wave_count":1},"fallback_work":3510372384,"features":{"max_depth":2,"packed_tokens":52699,"segment_count":14,"segments_below":[0,1,2,4,4,4,4,14]},"label":"minimum_effective_span_5303","median_ms":6188.192138671875,"n":8,"spread_pct":0.4850625334484561},{"cp_plan":{"max_rank_tokens":16896,"summary_ms":12.08455115556717,"wave_count":2},"fallback_work":3853549728,"features":{"max_depth":2,"packed_tokens":57931,"segment_count":14,"segments_below":[0,0,3,3,3,3,3,14]},"label":"minimum_effective_span_5395","median_ms":6486.96142578125,"n":8,"spread_pct":0.6888969061060509},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":7.619157433509827,"wave_count":1},"fallback_work":4201548224,"features":{"max_depth":2,"packed_tokens":63240,"segment_count":13,"segments_below":[1,1,1,1,1,1,1,13]},"label":"minimum_effective_span_5476","median_ms":6412.34375,"n":8,"spread_pct":0.28770651087989474},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.617275506258011,"wave_count":2},"fallback_work":586881792,"features":{"max_depth":8,"packed_tokens":6558,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_71","median_ms":2398.0257568359375,"n":8,"spread_pct":4.352426747527454},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.607567399740219,"wave_count":2},"fallback_work":574991488,"features":{"max_depth":7,"packed_tokens":6639,"segment_count":18,"segments_below":[6,14,16,17,17,17,17,18]},"label":"minimum_effective_span_82","median_ms":2199.83447265625,"n":8,"spread_pct":3.3094309075545234},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6494117230176926,"wave_count":2},"fallback_work":576172256,"features":{"max_depth":7,"packed_tokens":6657,"segment_count":18,"segments_below":[6,14,16,17,17,17,17,18]},"label":"minimum_effective_span_87","median_ms":2193.40283203125,"n":8,"spread_pct":31.898533176880512},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.631262108683586,"wave_count":2},"fallback_work":576893600,"features":{"max_depth":7,"packed_tokens":6668,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_93","median_ms":2195.8372802734375,"n":8,"spread_pct":4.142937584037268},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.6084260791540146,"wave_count":2},"fallback_work":575385152,"features":{"max_depth":7,"packed_tokens":6645,"segment_count":18,"segments_below":[7,12,17,17,17,17,17,18]},"label":"minimum_effective_span_94","median_ms":2211.9915771484375,"n":8,"spread_pct":3.5030891577475067},{"cp_plan":{"max_rank_tokens":18944,"summary_ms":8.903158828616142,"wave_count":1},"fallback_work":4521024608,"features":{"max_depth":1,"packed_tokens":68881,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":7703.608154296875,"n":8,"spread_pct":0.23305460144978862},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.8505960255861282,"wave_count":2},"fallback_work":642787360,"features":{"max_depth":11,"packed_tokens":6620,"segment_count":23,"segments_below":[16,19,20,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":2991.433349609375,"n":8,"spread_pct":4.663789690466209},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.954954445362091,"wave_count":2},"fallback_work":649364768,"features":{"max_depth":12,"packed_tokens":6458,"segment_count":24,"segments_below":[16,20,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":3180.415283203125,"n":8,"spread_pct":2.531269524381743},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.895508125424385,"wave_count":2},"fallback_work":655450208,"features":{"max_depth":13,"packed_tokens":6292,"segment_count":24,"segments_below":[17,21,23,23,23,23,23,24]},"label":"uniform_depth_12","median_ms":3371.5869140625,"n":8,"spread_pct":3.9336945203555103},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":5.521848797798157,"wave_count":1},"fallback_work":661533088,"features":{"max_depth":3,"packed_tokens":9004,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1952.8949584960938,"n":8,"spread_pct":1.0070498615915107},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":5.453107878565788,"wave_count":1},"fallback_work":643316608,"features":{"max_depth":4,"packed_tokens":8464,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1995.9979248046875,"n":8,"spread_pct":3.168596339782701},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":5.474867299199104,"wave_count":2},"fallback_work":637497184,"features":{"max_depth":5,"packed_tokens":8113,"segment_count":17,"segments_below":[8,9,11,13,16,16,16,17]},"label":"uniform_depth_4","median_ms":2087.3917236328125,"n":8,"spread_pct":3.729609618481685},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":5.087953060865402,"wave_count":2},"fallback_work":638433568,"features":{"max_depth":6,"packed_tokens":7865,"segment_count":18,"segments_below":[10,11,13,14,17,17,17,18]},"label":"uniform_depth_5","median_ms":2089.2587890625,"n":8,"spread_pct":3.4509893189729275},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.8046371191740036,"wave_count":2},"fallback_work":641403136,"features":{"max_depth":7,"packed_tokens":7648,"segment_count":19,"segments_below":[11,13,14,16,18,18,18,19]},"label":"uniform_depth_6","median_ms":2266.2933349609375,"n":8,"spread_pct":2.8941247126723124},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.8375332951545715,"wave_count":2},"fallback_work":638928736,"features":{"max_depth":8,"packed_tokens":7348,"segment_count":20,"segments_below":[14,14,15,18,19,19,19,20]},"label":"uniform_depth_7","median_ms":2460.2850341796875,"n":8,"spread_pct":4.996513212456378},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.832932561635971,"wave_count":2},"fallback_work":644325376,"features":{"max_depth":9,"packed_tokens":7168,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":2426.2095947265625,"n":8,"spread_pct":4.5274110967422265},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.9225518703460693,"wave_count":2},"fallback_work":643687552,"features":{"max_depth":10,"packed_tokens":6896,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":2566.341064453125,"n":8,"spread_pct":29.215039993881124}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"22fe33a263cacb6bc86ab088fe07af0a0fb8e68bfdc85c518a57ec37e403d138","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":6,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.309000447392464,"wave_count":2},"fallback_work":883905760,"features":{"max_depth":2,"packed_tokens":12674,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1926.984130859375,"n":8,"spread_pct":0.8887894186782673},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":4.340961575508118,"wave_count":1},"fallback_work":1648380576,"features":{"max_depth":2,"packed_tokens":24329,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11722","median_ms":4459.409912109375,"n":8,"spread_pct":8.93647742662316},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":8.349066600203514,"wave_count":2},"fallback_work":2417249888,"features":{"max_depth":2,"packed_tokens":36051,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11733","median_ms":4858.80126953125,"n":8,"spread_pct":0.15470074089334301},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":6.11887127161026,"wave_count":1},"fallback_work":3180872064,"features":{"max_depth":2,"packed_tokens":47693,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11735","median_ms":5357.909423828125,"n":8,"spread_pct":0.31782561530469544},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":11.96950487792492,"wave_count":3},"fallback_work":3942198240,"features":{"max_depth":2,"packed_tokens":59300,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11758","median_ms":6827.1533203125,"n":8,"spread_pct":0.38307081742896765},{"cp_plan":{"max_rank_tokens":20480,"summary_ms":8.138500154018402,"wave_count":1},"fallback_work":4710477504,"features":{"max_depth":2,"packed_tokens":71013,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11808","median_ms":9249.31982421875,"n":8,"spread_pct":0.35854621277842064},{"cp_plan":{"max_rank_tokens":22016,"summary_ms":9.480224922299385,"wave_count":1},"fallback_work":5487677056,"features":{"max_depth":2,"packed_tokens":82862,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11855","median_ms":9646.123046875,"n":8,"spread_pct":0.2380226835789557},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7506259977817535,"wave_count":2},"fallback_work":926416224,"features":{"max_depth":6,"packed_tokens":12273,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":4085.0506591796875,"n":8,"spread_pct":0.3978576253479969},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.6861933767795563,"wave_count":2},"fallback_work":892057152,"features":{"max_depth":3,"packed_tokens":12536,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":4080.8548583984375,"n":8,"spread_pct":16.450877305759438},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.722589462995529,"wave_count":2},"fallback_work":910065664,"features":{"max_depth":5,"packed_tokens":12286,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":4084.8583984375,"n":8,"spread_pct":0.3111122798922756},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.8479380309581757,"wave_count":2},"fallback_work":942701024,"features":{"max_depth":7,"packed_tokens":12259,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":4088.0601806640625,"n":8,"spread_pct":0.4931054430404807},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7017595022916794,"wave_count":2},"fallback_work":895223776,"features":{"max_depth":4,"packed_tokens":12322,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":4087.5653076171875,"n":8,"spread_pct":16.365516426817887},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7210285663604736,"wave_count":2},"fallback_work":897978528,"features":{"max_depth":4,"packed_tokens":12364,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":4085.085205078125,"n":8,"spread_pct":0.3409112166023637},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.6506131291389465,"wave_count":2},"fallback_work":884186112,"features":{"max_depth":3,"packed_tokens":12416,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":4084.9849853515625,"n":8,"spread_pct":0.30498864204032744},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.759589046239853,"wave_count":2},"fallback_work":925825728,"features":{"max_depth":6,"packed_tokens":12264,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":4088.182861328125,"n":8,"spread_pct":0.4182035003828262},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.6697499454021454,"wave_count":2},"fallback_work":883661376,"features":{"max_depth":3,"packed_tokens":12408,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":4084.2607421875,"n":8,"spread_pct":16.596467766528473},{"cp_plan":{"max_rank_tokens":25088,"summary_ms":10.54646447300911,"wave_count":1},"fallback_work":6214775008,"features":{"max_depth":1,"packed_tokens":94721,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":10048.88330078125,"n":8,"spread_pct":0.22123299322992057},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.905488342046738,"wave_count":2},"fallback_work":896779776,"features":{"max_depth":3,"packed_tokens":12608,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":4092.0753173828125,"n":8,"spread_pct":0.36310774914153365},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":4.81070950627327,"wave_count":2},"fallback_work":913327168,"features":{"max_depth":4,"packed_tokens":12598,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":4091.6314697265625,"n":8,"spread_pct":0.43541736248586427},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.77732515335083,"wave_count":2},"fallback_work":924495904,"features":{"max_depth":5,"packed_tokens":12506,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":4093.9959716796875,"n":8,"spread_pct":0.2908102939306336},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.7655457854270935,"wave_count":2},"fallback_work":931860192,"features":{"max_depth":6,"packed_tokens":12356,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":4092.7431640625,"n":8,"spread_pct":0.2756462765532267},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.8438495248556137,"wave_count":2},"fallback_work":942897856,"features":{"max_depth":7,"packed_tokens":12262,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":4093.0916748046875,"n":8,"spread_pct":0.4322081016904943},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":3.8079172372817993,"wave_count":2},"fallback_work":959773152,"features":{"max_depth":8,"packed_tokens":12257,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":4092.3175048828125,"n":8,"spread_pct":0.37124154698097034}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"d04b77db2ed7d695d20c6f183215ead28d0337ae88b9ddbc3674926310d364b9","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":7,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12659,"summary_ms":0.00039301812648773193,"wave_count":0},"fallback_work":882325888,"features":{"max_depth":2,"packed_tokens":12659,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":3204.6771240234375,"n":8,"spread_pct":0.5359978958366782},{"cp_plan":{"max_rank_tokens":24836,"summary_ms":0.0003166496753692627,"wave_count":0},"fallback_work":1681916416,"features":{"max_depth":2,"packed_tokens":24836,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12268","median_ms":6094.648193359375,"n":8,"spread_pct":0.18663114519298218},{"cp_plan":{"max_rank_tokens":37143,"summary_ms":0.0002998858690261841,"wave_count":0},"fallback_work":2490043264,"features":{"max_depth":2,"packed_tokens":37143,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12313","median_ms":8945.17431640625,"n":8,"spread_pct":0.15982779646650655},{"cp_plan":{"max_rank_tokens":12609,"summary_ms":0.00037066638469696045,"wave_count":0},"fallback_work":896049280,"features":{"max_depth":3,"packed_tokens":12609,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":3378.072265625,"n":8,"spread_pct":1.0704159350313631},{"cp_plan":{"max_rank_tokens":49460,"summary_ms":0.0005252659320831299,"wave_count":0},"fallback_work":3248265728,"features":{"max_depth":1,"packed_tokens":49460,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":11462.29931640625,"n":8,"spread_pct":0.241314203079738},{"cp_plan":{"max_rank_tokens":12569,"summary_ms":0.0003390014171600342,"wave_count":0},"fallback_work":893422720,"features":{"max_depth":3,"packed_tokens":12569,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":3369.0572509765625,"n":8,"spread_pct":0.36529164834042815},{"cp_plan":{"max_rank_tokens":12564,"summary_ms":0.000346451997756958,"wave_count":0},"fallback_work":910100992,"features":{"max_depth":4,"packed_tokens":12564,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":3486.4176025390625,"n":8,"spread_pct":0.5719461051647379}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"f8b37d8f36ed38a4d2ea74f20bfe0d15bedbdb66ffaae41eb728a1756d67d5ff","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":0,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12382,"summary_ms":0.00047124922275543213,"wave_count":0},"fallback_work":863874816,"features":{"max_depth":2,"packed_tokens":12382,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":3077.69189453125,"n":8,"spread_pct":0.6416511967975189},{"cp_plan":{"max_rank_tokens":24645,"summary_ms":0.0005383044481277466,"wave_count":0},"fallback_work":1618551424,"features":{"max_depth":1,"packed_tokens":24645,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5834.961181640625,"n":8,"spread_pct":0.5180501875292809}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"515df421b93fe5f98e250ae3d849975dccadfd7dc00046242b5875ddf392c74f","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":1,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12546,"summary_ms":0.00048801302909851074,"wave_count":0},"fallback_work":823951616,"features":{"max_depth":1,"packed_tokens":12546,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3082.0472412109375,"n":8,"spread_pct":22.044831754081123}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"8aa85999c5d3c576d127bd2e3a44436cbd0e9062c7c92b16bcbaab0814703094","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":2,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":47253,"summary_ms":0.0004284083843231201,"wave_count":0},"fallback_work":3154037376,"features":{"max_depth":2,"packed_tokens":47253,"segment_count":6,"segments_below":[0,0,0,1,1,1,1,3]},"label":"depth_one","median_ms":10886.3486328125,"n":8,"spread_pct":0.19293802686689832},{"cp_plan":{"max_rank_tokens":23150,"summary_ms":0.0003632158041000366,"wave_count":0},"fallback_work":1571567360,"features":{"max_depth":2,"packed_tokens":23150,"segment_count":7,"segments_below":[0,0,0,1,3,5,5,6]},"label":"minimum_effective_span_1366","median_ms":5536.30908203125,"n":8,"spread_pct":0.6188813356710022},{"cp_plan":{"max_rank_tokens":21785,"summary_ms":0.0003762543201446533,"wave_count":0},"fallback_work":1498811520,"features":{"max_depth":3,"packed_tokens":21785,"segment_count":7,"segments_below":[0,0,0,1,4,5,5,6]},"label":"minimum_effective_span_305","median_ms":5396.025146484375,"n":8,"spread_pct":0.46949344540222593},{"cp_plan":{"max_rank_tokens":28799,"summary_ms":0.00031478703022003174,"wave_count":0},"fallback_work":1942372224,"features":{"max_depth":2,"packed_tokens":28799,"segment_count":6,"segments_below":[0,0,0,1,3,3,3,4]},"label":"minimum_effective_span_7015","median_ms":6869.6796875,"n":8,"spread_pct":0.45535948842229046},{"cp_plan":{"max_rank_tokens":37178,"summary_ms":0.0003725290298461914,"wave_count":0},"fallback_work":2492472576,"features":{"max_depth":2,"packed_tokens":37178,"segment_count":6,"segments_below":[0,0,0,1,2,2,2,3]},"label":"minimum_effective_span_8380","median_ms":8698.81201171875,"n":8,"spread_pct":0.26964664434523683},{"cp_plan":{"max_rank_tokens":48469,"summary_ms":0.0005029141902923584,"wave_count":0},"fallback_work":3183323776,"features":{"max_depth":1,"packed_tokens":48469,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":10963.21875,"n":8,"spread_pct":0.23049372475578853},{"cp_plan":{"max_rank_tokens":22846,"summary_ms":0.0003725290298461914,"wave_count":0},"fallback_work":1568612096,"features":{"max_depth":3,"packed_tokens":22846,"segment_count":8,"segments_below":[0,0,0,2,4,6,6,7]},"label":"uniform_depth_2","median_ms":5529.7490234375,"n":8,"spread_pct":0.3856978844718279},{"cp_plan":{"max_rank_tokens":21481,"summary_ms":0.0003241002559661865,"wave_count":0},"fallback_work":1495856256,"features":{"max_depth":4,"packed_tokens":21481,"segment_count":8,"segments_below":[0,0,0,2,5,6,6,7]},"label":"uniform_depth_3","median_ms":5394.402099609375,"n":8,"spread_pct":0.4072780344895115}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"62112dd684879b09adbef34aa98e5ea3966b7de9b9e8067de5920b1d73f07197","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":3,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":30477,"summary_ms":0.0003129243850708008,"wave_count":0},"fallback_work":2052195968,"features":{"max_depth":2,"packed_tokens":30477,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,2]},"label":"depth_one","median_ms":7206.847900390625,"n":8,"spread_pct":0.3160649513813851},{"cp_plan":{"max_rank_tokens":19821,"summary_ms":0.0003688037395477295,"wave_count":0},"fallback_work":1352480384,"features":{"max_depth":2,"packed_tokens":19821,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_495","median_ms":4847.507568359375,"n":8,"spread_pct":0.38135738550804665},{"cp_plan":{"max_rank_tokens":31465,"summary_ms":0.00037439167499542236,"wave_count":0},"fallback_work":2066510976,"features":{"max_depth":1,"packed_tokens":31465,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":7248.92578125,"n":8,"spread_pct":0.27921701760767353},{"cp_plan":{"max_rank_tokens":19327,"summary_ms":0.00031851232051849365,"wave_count":0},"fallback_work":1337048960,"features":{"max_depth":3,"packed_tokens":19327,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,4]},"label":"uniform_depth_2","median_ms":4778.66845703125,"n":8,"spread_pct":0.178599290691369}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"28b4492a5c1ce50958108e27b72b7afdf09941e4670076dfa879b517d9cc7575","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":4,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13557,"summary_ms":0.00042282044887542725,"wave_count":0},"fallback_work":941423232,"features":{"max_depth":2,"packed_tokens":13557,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":3452.318603515625,"n":8,"spread_pct":1.3043345741726875},{"cp_plan":{"max_rank_tokens":25902,"summary_ms":0.0004079192876815796,"wave_count":0},"fallback_work":1752045312,"features":{"max_depth":2,"packed_tokens":25902,"segment_count":6,"segments_below":[0,1,2,4,4,4,4,4]},"label":"minimum_effective_span_12406","median_ms":6345.1640625,"n":8,"spread_pct":0.3060817108501676},{"cp_plan":{"max_rank_tokens":37995,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":2546120064,"features":{"max_depth":2,"packed_tokens":37995,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12426","median_ms":9222.2177734375,"n":8,"spread_pct":0.3571326082741545},{"cp_plan":{"max_rank_tokens":50582,"summary_ms":0.00028312206268310547,"wave_count":0},"fallback_work":3372632832,"features":{"max_depth":2,"packed_tokens":50582,"segment_count":6,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12592","median_ms":12172.2734375,"n":8,"spread_pct":0.1674447755561275},{"cp_plan":{"max_rank_tokens":13367,"summary_ms":0.00031478703022003174,"wave_count":0},"fallback_work":945953664,"features":{"max_depth":3,"packed_tokens":13367,"segment_count":7,"segments_below":[1,3,5,6,6,6,6,6]},"label":"minimum_effective_span_187","median_ms":3584.3568115234375,"n":8,"spread_pct":0.7178068791284665},{"cp_plan":{"max_rank_tokens":13185,"summary_ms":0.00025890767574310303,"wave_count":0},"fallback_work":934002816,"features":{"max_depth":3,"packed_tokens":13185,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_21","median_ms":3552.011962890625,"n":8,"spread_pct":0.7121500310570298},{"cp_plan":{"max_rank_tokens":63177,"summary_ms":0.00036135315895080566,"wave_count":0},"fallback_work":4149109888,"features":{"max_depth":1,"packed_tokens":63177,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":14645.142578125,"n":8,"spread_pct":0.2325389724328659},{"cp_plan":{"max_rank_tokens":13497,"summary_ms":0.00033155083656311035,"wave_count":0},"fallback_work":954489984,"features":{"max_depth":3,"packed_tokens":13497,"segment_count":7,"segments_below":[1,2,4,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":3620.05859375,"n":8,"spread_pct":0.6299198260559923},{"cp_plan":{"max_rank_tokens":13165,"summary_ms":0.00027008354663848877,"wave_count":0},"fallback_work":949696128,"features":{"max_depth":4,"packed_tokens":13165,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":3695.4793701171875,"n":8,"spread_pct":0.6777310234438079},{"cp_plan":{"max_rank_tokens":13161,"summary_ms":0.0005923211574554443,"wave_count":0},"fallback_work":966440064,"features":{"max_depth":5,"packed_tokens":13161,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":3837.2249755859375,"n":8,"spread_pct":0.4115217566201924}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"7241823ac1b67d127c5921d771dda170dcc9d4886feffaff3c2520113cd01a08","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":5,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9433,"summary_ms":0.00046938657760620117,"wave_count":0},"fallback_work":671673472,"features":{"max_depth":2,"packed_tokens":9433,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":2338.3431396484375,"n":8,"spread_pct":0.7346009207595418},{"cp_plan":{"max_rank_tokens":6738,"summary_ms":0.0004433095455169678,"wave_count":0},"fallback_work":562604288,"features":{"max_depth":6,"packed_tokens":6738,"segment_count":17,"segments_below":[6,11,15,16,16,16,16,17]},"label":"minimum_effective_span_105","median_ms":2401.5234375,"n":8,"spread_pct":3.3621033198327885},{"cp_plan":{"max_rank_tokens":6778,"summary_ms":0.00033527612686157227,"wave_count":0},"fallback_work":565230848,"features":{"max_depth":6,"packed_tokens":6778,"segment_count":17,"segments_below":[5,10,15,16,16,16,16,17]},"label":"minimum_effective_span_113","median_ms":2394.5791015625,"n":8,"spread_pct":3.4023280122053023},{"cp_plan":{"max_rank_tokens":6877,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":571731584,"features":{"max_depth":6,"packed_tokens":6877,"segment_count":17,"segments_below":[5,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":2470.25341796875,"n":8,"spread_pct":2.8448261037236224},{"cp_plan":{"max_rank_tokens":7090,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":568711424,"features":{"max_depth":5,"packed_tokens":7090,"segment_count":16,"segments_below":[3,8,14,15,15,15,15,16]},"label":"minimum_effective_span_164","median_ms":2307.3062744140625,"n":8,"spread_pct":1.9414909531029176},{"cp_plan":{"max_rank_tokens":7134,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":571600640,"features":{"max_depth":5,"packed_tokens":7134,"segment_count":16,"segments_below":[4,7,14,15,15,15,15,16]},"label":"minimum_effective_span_195","median_ms":2305.4337158203125,"n":8,"spread_pct":2.2084163109688055},{"cp_plan":{"max_rank_tokens":7415,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":573176704,"features":{"max_depth":4,"packed_tokens":7415,"segment_count":16,"segments_below":[3,5,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":2255.445068359375,"n":8,"spread_pct":3.855350923216093},{"cp_plan":{"max_rank_tokens":7626,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":586900736,"features":{"max_depth":4,"packed_tokens":7626,"segment_count":15,"segments_below":[3,6,10,14,14,14,14,15]},"label":"minimum_effective_span_245","median_ms":2251.0753173828125,"n":8,"spread_pct":2.0776962659284632},{"cp_plan":{"max_rank_tokens":7702,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":591891200,"features":{"max_depth":4,"packed_tokens":7702,"segment_count":15,"segments_below":[3,5,9,14,14,14,14,15]},"label":"minimum_effective_span_281","median_ms":2273.771484375,"n":8,"spread_pct":2.0103468476710034},{"cp_plan":{"max_rank_tokens":6323,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":637393280,"features":{"max_depth":12,"packed_tokens":6323,"segment_count":23,"segments_below":[16,20,22,22,22,22,22,23]},"label":"minimum_effective_span_32","median_ms":3649.9779052734375,"n":8,"spread_pct":3.64155671827396},{"cp_plan":{"max_rank_tokens":8041,"summary_ms":0.00019371509552001953,"wave_count":0},"fallback_work":597275776,"features":{"max_depth":3,"packed_tokens":8041,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_340","median_ms":2235.3056640625,"n":8,"spread_pct":1.0958394514872083},{"cp_plan":{"max_rank_tokens":8113,"summary_ms":0.0004153698682785034,"wave_count":0},"fallback_work":602003584,"features":{"max_depth":3,"packed_tokens":8113,"segment_count":15,"segments_below":[1,4,8,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":2253.9873046875,"n":8,"spread_pct":1.841376588276938},{"cp_plan":{"max_rank_tokens":6359,"summary_ms":0.0003296881914138794,"wave_count":0},"fallback_work":622750592,"features":{"max_depth":11,"packed_tokens":6359,"segment_count":22,"segments_below":[15,19,21,21,21,21,21,22]},"label":"minimum_effective_span_37","median_ms":3443.2213134765625,"n":8,"spread_pct":2.919749795696928},{"cp_plan":{"max_rank_tokens":6437,"summary_ms":0.00021606683731079102,"wave_count":0},"fallback_work":593859200,"features":{"max_depth":9,"packed_tokens":6437,"segment_count":20,"segments_below":[9,17,19,19,19,19,19,20]},"label":"minimum_effective_span_40","median_ms":3042.498291015625,"n":8,"spread_pct":4.717553813077488},{"cp_plan":{"max_rank_tokens":8391,"summary_ms":0.00020116567611694336,"wave_count":0},"fallback_work":620258176,"features":{"max_depth":3,"packed_tokens":8391,"segment_count":15,"segments_below":[1,3,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":2306.3389892578125,"n":8,"spread_pct":1.906568116522858},{"cp_plan":{"max_rank_tokens":13958,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":968803072,"features":{"max_depth":2,"packed_tokens":13958,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4955","median_ms":3305.878173828125,"n":8,"spread_pct":0.4594089437486213},{"cp_plan":{"max_rank_tokens":18411,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":1261204864,"features":{"max_depth":2,"packed_tokens":18411,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4994","median_ms":4298.830322265625,"n":8,"spread_pct":0.6953653874211445},{"cp_plan":{"max_rank_tokens":23107,"summary_ms":0.00020116567611694336,"wave_count":0},"fallback_work":1569563008,"features":{"max_depth":2,"packed_tokens":23107,"segment_count":14,"segments_below":[2,3,5,7,10,10,10,14]},"label":"minimum_effective_span_5048","median_ms":5296.017822265625,"n":8,"spread_pct":0.3945697517698407},{"cp_plan":{"max_rank_tokens":27945,"summary_ms":0.00020302832126617432,"wave_count":0},"fallback_work":1887245440,"features":{"max_depth":2,"packed_tokens":27945,"segment_count":14,"segments_below":[2,3,5,6,9,9,9,14]},"label":"minimum_effective_span_5087","median_ms":6301.966552734375,"n":8,"spread_pct":0.5210582091752346},{"cp_plan":{"max_rank_tokens":32845,"summary_ms":0.00021792948246002197,"wave_count":0},"fallback_work":2208999040,"features":{"max_depth":2,"packed_tokens":32845,"segment_count":14,"segments_below":[1,3,4,6,8,8,8,14]},"label":"minimum_effective_span_5118","median_ms":7370.2685546875,"n":8,"spread_pct":0.9172661751348686},{"cp_plan":{"max_rank_tokens":37693,"summary_ms":0.0004675239324569702,"wave_count":0},"fallback_work":2527338112,"features":{"max_depth":2,"packed_tokens":37693,"segment_count":14,"segments_below":[2,2,3,6,7,7,7,14]},"label":"minimum_effective_span_5149","median_ms":8372.2646484375,"n":8,"spread_pct":8.750712540324795},{"cp_plan":{"max_rank_tokens":42711,"summary_ms":0.0003259629011154175,"wave_count":0},"fallback_work":2856840064,"features":{"max_depth":2,"packed_tokens":42711,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_5199","median_ms":9440.51953125,"n":8,"spread_pct":0.19354320823676863},{"cp_plan":{"max_rank_tokens":8746,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":643437824,"features":{"max_depth":3,"packed_tokens":8746,"segment_count":14,"segments_below":[2,4,7,10,13,13,13,14]},"label":"minimum_effective_span_522","median_ms":2181.3719482421875,"n":8,"spread_pct":0.5728995608736281},{"cp_plan":{"max_rank_tokens":47673,"summary_ms":0.0002253800630569458,"wave_count":0},"fallback_work":3182664832,"features":{"max_depth":2,"packed_tokens":47673,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5235","median_ms":10557.208984375,"n":8,"spread_pct":0.23249442074441506},{"cp_plan":{"max_rank_tokens":52699,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":3512692096,"features":{"max_depth":2,"packed_tokens":52699,"segment_count":14,"segments_below":[0,1,2,4,4,4,4,14]},"label":"minimum_effective_span_5303","median_ms":11556.115234375,"n":8,"spread_pct":0.20332173289608738},{"cp_plan":{"max_rank_tokens":57931,"summary_ms":0.00021420419216156006,"wave_count":0},"fallback_work":3856246144,"features":{"max_depth":2,"packed_tokens":57931,"segment_count":14,"segments_below":[0,0,3,3,3,3,3,14]},"label":"minimum_effective_span_5395","median_ms":12769.83154296875,"n":8,"spread_pct":0.23470693083655303},{"cp_plan":{"max_rank_tokens":63240,"summary_ms":0.0002980232238769531,"wave_count":0},"fallback_work":4204725248,"features":{"max_depth":2,"packed_tokens":63240,"segment_count":13,"segments_below":[1,1,1,1,1,1,1,13]},"label":"minimum_effective_span_5476","median_ms":13735.08154296875,"n":8,"spread_pct":0.1393628534538912},{"cp_plan":{"max_rank_tokens":6558,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":584797952,"features":{"max_depth":8,"packed_tokens":6558,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_71","median_ms":2756.5318603515625,"n":8,"spread_pct":3.947380133010597},{"cp_plan":{"max_rank_tokens":6639,"summary_ms":0.0002123415470123291,"wave_count":0},"fallback_work":573110144,"features":{"max_depth":7,"packed_tokens":6639,"segment_count":18,"segments_below":[6,14,16,17,17,17,17,18]},"label":"minimum_effective_span_82","median_ms":2624.5767822265625,"n":8,"spread_pct":4.13855786737938},{"cp_plan":{"max_rank_tokens":6657,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":574292096,"features":{"max_depth":7,"packed_tokens":6657,"segment_count":18,"segments_below":[6,14,16,17,17,17,17,18]},"label":"minimum_effective_span_87","median_ms":2668.0157470703125,"n":8,"spread_pct":3.500377578451331},{"cp_plan":{"max_rank_tokens":6668,"summary_ms":0.00021047890186309814,"wave_count":0},"fallback_work":575014400,"features":{"max_depth":7,"packed_tokens":6668,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_93","median_ms":2568.2969970703125,"n":8,"spread_pct":5.67178223767697},{"cp_plan":{"max_rank_tokens":6645,"summary_ms":0.0002738088369369507,"wave_count":0},"fallback_work":573504128,"features":{"max_depth":7,"packed_tokens":6645,"segment_count":18,"segments_below":[7,12,17,17,17,17,17,18]},"label":"minimum_effective_span_94","median_ms":2651.528076171875,"n":8,"spread_pct":3.325746623662523},{"cp_plan":{"max_rank_tokens":68881,"summary_ms":0.0005904585123062134,"wave_count":0},"fallback_work":4524705920,"features":{"max_depth":1,"packed_tokens":68881,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":14865.19775390625,"n":8,"spread_pct":0.6700253006141619},{"cp_plan":{"max_rank_tokens":6620,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":640019968,"features":{"max_depth":11,"packed_tokens":6620,"segment_count":23,"segments_below":[16,19,20,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":3408.2330322265625,"n":8,"spread_pct":21.76914278510092},{"cp_plan":{"max_rank_tokens":6458,"summary_ms":0.0001955777406692505,"wave_count":0},"fallback_work":646388992,"features":{"max_depth":12,"packed_tokens":6458,"segment_count":24,"segments_below":[16,20,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":3594.5982666015625,"n":8,"spread_pct":2.572117063808546},{"cp_plan":{"max_rank_tokens":6292,"summary_ms":0.00028312206268310547,"wave_count":0},"fallback_work":652364288,"features":{"max_depth":13,"packed_tokens":6292,"segment_count":24,"segments_below":[17,21,23,23,23,23,23,24]},"label":"uniform_depth_12","median_ms":3716.8345947265625,"n":8,"spread_pct":3.7692996349977927},{"cp_plan":{"max_rank_tokens":9004,"summary_ms":0.00030547380447387695,"wave_count":0},"fallback_work":660510208,"features":{"max_depth":3,"packed_tokens":9004,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":2405.82958984375,"n":8,"spread_pct":1.2538919644848663},{"cp_plan":{"max_rank_tokens":8464,"summary_ms":0.0002253800630569458,"wave_count":0},"fallback_work":642058240,"features":{"max_depth":4,"packed_tokens":8464,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":2440.45361328125,"n":8,"spread_pct":1.1809408031013702},{"cp_plan":{"max_rank_tokens":8113,"summary_ms":0.0003203749656677246,"wave_count":0},"fallback_work":636016768,"features":{"max_depth":5,"packed_tokens":8113,"segment_count":17,"segments_below":[8,9,11,13,16,16,16,17]},"label":"uniform_depth_4","median_ms":2532.37646484375,"n":8,"spread_pct":1.5992015127527162},{"cp_plan":{"max_rank_tokens":7865,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":636738688,"features":{"max_depth":6,"packed_tokens":7865,"segment_count":18,"segments_below":[10,11,13,14,17,17,17,18]},"label":"uniform_depth_5","median_ms":2627.1375732421875,"n":8,"spread_pct":4.519264469859946},{"cp_plan":{"max_rank_tokens":7648,"summary_ms":0.0002868473529815674,"wave_count":0},"fallback_work":639496192,"features":{"max_depth":7,"packed_tokens":7648,"segment_count":19,"segments_below":[11,13,14,16,18,18,18,19]},"label":"uniform_depth_6","median_ms":2712.4482421875,"n":8,"spread_pct":1.814838776805577},{"cp_plan":{"max_rank_tokens":7348,"summary_ms":0.00021047890186309814,"wave_count":0},"fallback_work":636803584,"features":{"max_depth":8,"packed_tokens":7348,"segment_count":20,"segments_below":[14,14,15,18,19,19,19,20]},"label":"uniform_depth_7","median_ms":2879.7034912109375,"n":8,"spread_pct":2.4080000573111304},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":0.00029616057872772217,"wave_count":0},"fallback_work":641990656,"features":{"max_depth":9,"packed_tokens":7168,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":3013.8475341796875,"n":8,"spread_pct":4.508923835814981},{"cp_plan":{"max_rank_tokens":6896,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":641136640,"features":{"max_depth":10,"packed_tokens":6896,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":3228.5135498046875,"n":8,"spread_pct":4.586164472353147}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"22fe33a263cacb6bc86ab088fe07af0a0fb8e68bfdc85c518a57ec37e403d138","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":6,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":12674,"summary_ms":0.0003594905138015747,"wave_count":0},"fallback_work":883835136,"features":{"max_depth":2,"packed_tokens":12674,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":3228.601318359375,"n":8,"spread_pct":0.45692223383967956},{"cp_plan":{"max_rank_tokens":24329,"summary_ms":0.00029616057872772217,"wave_count":0},"fallback_work":1649149056,"features":{"max_depth":2,"packed_tokens":24329,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11722","median_ms":6050.924072265625,"n":8,"spread_pct":0.7735695392922657},{"cp_plan":{"max_rank_tokens":36051,"summary_ms":0.00024400651454925537,"wave_count":0},"fallback_work":2418862464,"features":{"max_depth":2,"packed_tokens":36051,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11733","median_ms":8705.775390625,"n":8,"spread_pct":0.34388091418300126},{"cp_plan":{"max_rank_tokens":47693,"summary_ms":0.000346451997756958,"wave_count":0},"fallback_work":3183322752,"features":{"max_depth":2,"packed_tokens":47693,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11735","median_ms":11450.84765625,"n":8,"spread_pct":0.42094829906492326},{"cp_plan":{"max_rank_tokens":59300,"summary_ms":0.0002514570951461792,"wave_count":0},"fallback_work":3945484800,"features":{"max_depth":2,"packed_tokens":59300,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11758","median_ms":14046.99267578125,"n":8,"spread_pct":0.40932638182501324},{"cp_plan":{"max_rank_tokens":71013,"summary_ms":0.00022724270820617676,"wave_count":0},"fallback_work":4714607232,"features":{"max_depth":2,"packed_tokens":71013,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11808","median_ms":16838.439453125,"n":8,"spread_pct":3.935389648011705},{"cp_plan":{"max_rank_tokens":82862,"summary_ms":0.00024028122425079346,"wave_count":0},"fallback_work":5492659968,"features":{"max_depth":2,"packed_tokens":82862,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11855","median_ms":19422.3642578125,"n":8,"spread_pct":0.3622193547919976},{"cp_plan":{"max_rank_tokens":12273,"summary_ms":0.00024028122425079346,"wave_count":0},"fallback_work":925530240,"features":{"max_depth":6,"packed_tokens":12273,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":3765.091064453125,"n":8,"spread_pct":0.5491442543104368},{"cp_plan":{"max_rank_tokens":12536,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":891780096,"features":{"max_depth":3,"packed_tokens":12536,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":3374.99365234375,"n":8,"spread_pct":1.1448950931186648},{"cp_plan":{"max_rank_tokens":12286,"summary_ms":0.000260770320892334,"wave_count":0},"fallback_work":909377280,"features":{"max_depth":5,"packed_tokens":12286,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":3591.8126220703125,"n":8,"spread_pct":0.8158338348774911},{"cp_plan":{"max_rank_tokens":12259,"summary_ms":0.00024586915969848633,"wave_count":0},"fallback_work":941617536,"features":{"max_depth":7,"packed_tokens":12259,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":3909.232421875,"n":8,"spread_pct":0.8062968923227116},{"cp_plan":{"max_rank_tokens":12322,"summary_ms":0.00029616057872772217,"wave_count":0},"fallback_work":894734592,"features":{"max_depth":4,"packed_tokens":12322,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":3469.0731201171875,"n":8,"spread_pct":0.8214887922004168},{"cp_plan":{"max_rank_tokens":12364,"summary_ms":0.0002719461917877197,"wave_count":0},"fallback_work":897492480,"features":{"max_depth":4,"packed_tokens":12364,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":3475.041748046875,"n":8,"spread_pct":0.7658195428323247},{"cp_plan":{"max_rank_tokens":12416,"summary_ms":0.00029243528842926025,"wave_count":0},"fallback_work":883900416,"features":{"max_depth":3,"packed_tokens":12416,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":3338.7528076171875,"n":8,"spread_pct":0.6424832555868121},{"cp_plan":{"max_rank_tokens":12264,"summary_ms":0.00026635825634002686,"wave_count":0},"fallback_work":924939264,"features":{"max_depth":6,"packed_tokens":12264,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":3728.4798583984375,"n":8,"spread_pct":18.520207779625437},{"cp_plan":{"max_rank_tokens":12408,"summary_ms":0.0002682209014892578,"wave_count":0},"fallback_work":883375104,"features":{"max_depth":3,"packed_tokens":12408,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":3336.0992431640625,"n":8,"spread_pct":0.9351786039355162},{"cp_plan":{"max_rank_tokens":94721,"summary_ms":0.0005234032869338989,"wave_count":0},"fallback_work":6220808320,"features":{"max_depth":1,"packed_tokens":94721,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":21755.119140625,"n":8,"spread_pct":0.11798587993971607},{"cp_plan":{"max_rank_tokens":12608,"summary_ms":0.00024028122425079346,"wave_count":0},"fallback_work":896507904,"features":{"max_depth":3,"packed_tokens":12608,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":3420.968994140625,"n":8,"spread_pct":0.8105742038292245},{"cp_plan":{"max_rank_tokens":12598,"summary_ms":0.000346451997756958,"wave_count":0},"fallback_work":912857856,"features":{"max_depth":4,"packed_tokens":12598,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":3527.753662109375,"n":8,"spread_pct":0.5501575502127072},{"cp_plan":{"max_rank_tokens":12506,"summary_ms":0.0003241002559661865,"wave_count":0},"fallback_work":923823360,"features":{"max_depth":5,"packed_tokens":12506,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":3658.9866943359375,"n":8,"spread_pct":0.5764982467216462},{"cp_plan":{"max_rank_tokens":12356,"summary_ms":0.00024400651454925537,"wave_count":0},"fallback_work":930980352,"features":{"max_depth":6,"packed_tokens":12356,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":3770.3157958984375,"n":8,"spread_pct":0.9755547315330453},{"cp_plan":{"max_rank_tokens":12262,"summary_ms":0.0003073364496231079,"wave_count":0},"fallback_work":941814528,"features":{"max_depth":7,"packed_tokens":12262,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":3878.9422607421875,"n":8,"spread_pct":6.67302465564346},{"cp_plan":{"max_rank_tokens":12257,"summary_ms":0.00030547380447387695,"wave_count":0},"fallback_work":958492800,"features":{"max_depth":8,"packed_tokens":12257,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":4042.9993896484375,"n":8,"spread_pct":2.2740467757928524}],"cell":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"d04b77db2ed7d695d20c6f183215ead28d0337ae88b9ddbc3674926310d364b9","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"group":7,"kind":"ellavox"}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":38879,"summary_ms":23.633582517504692,"wave_count":1},"fallback_work":4958631424,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":14524.69482421875,"n":8,"spread_pct":0.2056169491781827},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":7.8278034925460815,"wave_count":1},"fallback_work":994604384,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":3096.8897705078125,"n":8,"spread_pct":0.7985109493320798},{"cp_plan":{"max_rank_tokens":55296,"summary_ms":9.127773344516754,"wave_count":1},"fallback_work":6988884992,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":19633.3623046875,"n":8,"spread_pct":0.34313603334725556},{"cp_plan":{"max_rank_tokens":8704,"summary_ms":7.95806385576725,"wave_count":1},"fallback_work":944551424,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":3065.43115234375,"n":8,"spread_pct":0.7716466526059322}],"cell":"cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,2,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19456,"summary_ms":103.24226878583431,"wave_count":1},"fallback_work":4959066368,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":7492.074462890625,"n":8,"spread_pct":0.3856484148597523},{"cp_plan":{"max_rank_tokens":4608,"summary_ms":6.563352420926094,"wave_count":3},"fallback_work":996620544,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1779.1236572265625,"n":8,"spread_pct":1.759680089602576},{"cp_plan":{"max_rank_tokens":26880,"summary_ms":13.763798400759697,"wave_count":1},"fallback_work":6988427776,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":9963.818359375,"n":8,"spread_pct":0.35699479072229107},{"cp_plan":{"max_rank_tokens":4096,"summary_ms":6.765209138393402,"wave_count":1},"fallback_work":946723072,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1730.640625,"n":8,"spread_pct":1.687678029721653}],"cell":"cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":74719,"summary_ms":0.00030547380447387695,"wave_count":0},"fallback_work":4961103744,"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":15922.04443359375,"n":8,"spread_pct":0.16893833962206056},{"cp_plan":{"max_rank_tokens":14302,"summary_ms":0.00020675361156463623,"wave_count":0},"fallback_work":994111232,"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":3302.720947265625,"n":8,"spread_pct":0.2888243558057044},{"cp_plan":{"max_rank_tokens":106432,"summary_ms":0.0006016343832015991,"wave_count":0},"fallback_work":6992945152,"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":22094.279296875,"n":8,"spread_pct":0.2527432378520635},{"cp_plan":{"max_rank_tokens":13279,"summary_ms":0.0004079192876815796,"wave_count":0},"fallback_work":943943552,"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":3131.4727783203125,"n":8,"spread_pct":0.6868117892529244}],"cell":"cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":58351,"summary_ms":0.00040978193283081055,"wave_count":0},"fallback_work":3884218240,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":21794.751953125,"n":8,"spread_pct":0.43378777126410073},{"cp_plan":{"max_rank_tokens":24556,"summary_ms":0.000433996319770813,"wave_count":0},"fallback_work":1665791488,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":9255.08837890625,"n":8,"spread_pct":0.46427163351494616},{"cp_plan":{"max_rank_tokens":73696,"summary_ms":0.0011324882507324219,"wave_count":0},"fallback_work":4841271296,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":26527.9921875,"n":8,"spread_pct":0.24045944393054153},{"cp_plan":{"max_rank_tokens":21487,"summary_ms":0.00035762786865234375,"wave_count":0},"fallback_work":1481275264,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":8293.32080078125,"n":8,"spread_pct":0.41651552592474145}],"cell":"cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":31727,"summary_ms":11.259926483035088,"wave_count":1},"fallback_work":3882007296,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":11719.85205078125,"n":8,"spread_pct":0.4128111785487506},{"cp_plan":{"max_rank_tokens":14848,"summary_ms":8.631093427538872,"wave_count":1},"fallback_work":1665399232,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":5057.566650390625,"n":8,"spread_pct":0.28718186168398185},{"cp_plan":{"max_rank_tokens":37888,"summary_ms":7.088415324687958,"wave_count":1},"fallback_work":4838258176,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":13823.6455078125,"n":8,"spread_pct":0.23451806033494044},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":10.289022698998451,"wave_count":1},"fallback_work":1481095936,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":4763.034912109375,"n":8,"spread_pct":0.3154173290186249}],"cell":"cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,2,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":15360,"summary_ms":9.207133203744888,"wave_count":1},"fallback_work":3881786496,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":6020.5732421875,"n":8,"spread_pct":0.5285100713422903},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":7.406298071146011,"wave_count":1},"fallback_work":1666382752,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":2552.083984375,"n":8,"spread_pct":0.6234469072839522},{"cp_plan":{"max_rank_tokens":18432,"summary_ms":8.424637839198112,"wave_count":1},"fallback_work":4837538048,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":7072.060302734375,"n":8,"spread_pct":0.3507144073051831},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":6.641468033194542,"wave_count":1},"fallback_work":1482284160,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":2574.411865234375,"n":8,"spread_pct":0.6428292414709784}],"cell":"cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":58351,"summary_ms":0.00046193599700927734,"wave_count":0},"fallback_work":3884218240,"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":12636.03125,"n":8,"spread_pct":0.28643753106815084},{"cp_plan":{"max_rank_tokens":24556,"summary_ms":0.0004116445779800415,"wave_count":0},"fallback_work":1665791488,"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":5431.5947265625,"n":8,"spread_pct":0.4443852725832096},{"cp_plan":{"max_rank_tokens":73696,"summary_ms":0.0005234032869338989,"wave_count":0},"fallback_work":4841271296,"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":15367.97705078125,"n":8,"spread_pct":0.17746927529810128},{"cp_plan":{"max_rank_tokens":21487,"summary_ms":0.0005029141902923584,"wave_count":0},"fallback_work":1481275264,"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":4932.406005859375,"n":8,"spread_pct":0.4763815120772098}],"cell":"cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":37871,"summary_ms":0.0012703239917755127,"wave_count":0},"fallback_work":2539419520,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":13882.169921875,"n":8,"spread_pct":0.36733551679586746},{"cp_plan":{"max_rank_tokens":10222,"summary_ms":0.0004898756742477417,"wave_count":0},"fallback_work":724104960,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":3833.0582275390625,"n":8,"spread_pct":0.6329853050021985},{"cp_plan":{"max_rank_tokens":53216,"summary_ms":0.0017043203115463257,"wave_count":0},"fallback_work":3496472576,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":18670.2880859375,"n":8,"spread_pct":0.37808652670532933},{"cp_plan":{"max_rank_tokens":9199,"summary_ms":0.0007301568984985352,"wave_count":0},"fallback_work":673937280,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":3519.516357421875,"n":8,"spread_pct":0.8045608406084405}],"cell":"cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":21487,"summary_ms":98.06468710303307,"wave_count":1},"fallback_work":2538191616,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":7686.331298828125,"n":8,"spread_pct":0.19365876030505527},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.355844438076019,"wave_count":1},"fallback_work":724269664,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":2211.5989990234375,"n":8,"spread_pct":0.5268410472816244},{"cp_plan":{"max_rank_tokens":28672,"summary_ms":16.72654040157795,"wave_count":1},"fallback_work":3494442496,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":9949.587890625,"n":8,"spread_pct":0.23664218190569686},{"cp_plan":{"max_rank_tokens":6144,"summary_ms":8.12111422419548,"wave_count":1},"fallback_work":674216704,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":2173.4249267578125,"n":8,"spread_pct":1.0459037859504152}],"cell":"cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,2,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":10240,"summary_ms":113.9500830322504,"wave_count":1},"fallback_work":2538462336,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3968.7742919921875,"n":8,"spread_pct":0.9082744603033464},{"cp_plan":{"max_rank_tokens":3584,"summary_ms":3.8220416754484177,"wave_count":3},"fallback_work":725335168,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1446.10791015625,"n":8,"spread_pct":1.9488430184874606},{"cp_plan":{"max_rank_tokens":13312,"summary_ms":12.233974412083626,"wave_count":1},"fallback_work":3494213888,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":5057.420166015625,"n":8,"spread_pct":0.7974919664035957},{"cp_plan":{"max_rank_tokens":3072,"summary_ms":3.9435848593711853,"wave_count":1},"fallback_work":675437696,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1396.7614135742188,"n":8,"spread_pct":4.718626842341023}],"cell":"cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":37871,"summary_ms":0.0008009374141693115,"wave_count":0},"fallback_work":2539419520,"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":8150.342529296875,"n":8,"spread_pct":0.4673404045668494},{"cp_plan":{"max_rank_tokens":10222,"summary_ms":0.0003818422555923462,"wave_count":0},"fallback_work":724104960,"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":2408.2083740234375,"n":8,"spread_pct":0.4064873540675068},{"cp_plan":{"max_rank_tokens":53216,"summary_ms":0.0017546117305755615,"wave_count":0},"fallback_work":3496472576,"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":10983.54248046875,"n":8,"spread_pct":0.6789365925893655},{"cp_plan":{"max_rank_tokens":9199,"summary_ms":0.00030919909477233887,"wave_count":0},"fallback_work":673937280,"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":2269.06103515625,"n":8,"spread_pct":0.8310456749107011}],"cell":"cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":28852,"summary_ms":0.00039674341678619385,"wave_count":0},"fallback_work":1947884032,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":10731.72607421875,"n":8,"spread_pct":0.4890489943046808},{"cp_plan":{"max_rank_tokens":33460,"summary_ms":0.0002514570951461792,"wave_count":0},"fallback_work":2250004992,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":12268.26513671875,"n":8,"spread_pct":0.4592005216074709},{"cp_plan":{"max_rank_tokens":29364,"summary_ms":0.0003501772880554199,"wave_count":0},"fallback_work":1981274624,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":10864.04541015625,"n":8,"spread_pct":0.1188337839413939},{"cp_plan":{"max_rank_tokens":36532,"summary_ms":0.00023469328880310059,"wave_count":0},"fallback_work":2451495424,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":13444.0302734375,"n":8,"spread_pct":0.17827818704674553},{"cp_plan":{"max_rank_tokens":48820,"summary_ms":0.0009890645742416382,"wave_count":0},"fallback_work":3207813632,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":17540.02734375,"n":8,"spread_pct":0.3891322682248883}],"cell":"cal-hetero2|Qwen/Qwen3.5-27B|L64|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":18432,"summary_ms":5.6169964373111725,"wave_count":1},"fallback_work":1947285568,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":5895.01025390625,"n":8,"spread_pct":0.9250739316163921},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":4.813315346837044,"wave_count":1},"fallback_work":2249054272,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":7171.86669921875,"n":8,"spread_pct":0.17882645969838065},{"cp_plan":{"max_rank_tokens":18944,"summary_ms":5.436144769191742,"wave_count":1},"fallback_work":1980586048,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":6331.3076171875,"n":8,"spread_pct":0.707090369210917},{"cp_plan":{"max_rank_tokens":25088,"summary_ms":6.184142082929611,"wave_count":1},"fallback_work":2450331712,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":7547.35400390625,"n":8,"spread_pct":0.3359129444488545},{"cp_plan":{"max_rank_tokens":30720,"summary_ms":40.34462384879589,"wave_count":1},"fallback_work":3205994560,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":9870.16357421875,"n":8,"spread_pct":0.24024820304841746}],"cell":"cal-hetero2|Qwen/Qwen3.5-27B|L64|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,2,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":13312,"summary_ms":17.96460896730423,"wave_count":3},"fallback_work":1948165984,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":3728.6370849609375,"n":8,"spread_pct":1.1543298664805806},{"cp_plan":{"max_rank_tokens":16384,"summary_ms":6.297225132584572,"wave_count":2},"fallback_work":2249561952,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":4086.435791015625,"n":8,"spread_pct":0.4116431546559122},{"cp_plan":{"max_rank_tokens":13824,"summary_ms":18.01932416856289,"wave_count":3},"fallback_work":1981323104,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":3856.0919189453125,"n":8,"spread_pct":0.47743043052343515},{"cp_plan":{"max_rank_tokens":17920,"summary_ms":6.507042795419693,"wave_count":2},"fallback_work":2450634592,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":4666.44140625,"n":8,"spread_pct":0.5739116313371153},{"cp_plan":{"max_rank_tokens":22528,"summary_ms":7.728882133960724,"wave_count":1},"fallback_work":3205871456,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":5540.6484375,"n":8,"spread_pct":0.34927411474570746}],"cell":"cal-hetero2|Qwen/Qwen3.5-27B|L64|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":28852,"summary_ms":0.0002998858690261841,"wave_count":0},"fallback_work":1947884032,"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":6324.8193359375,"n":8,"spread_pct":0.7394063102407662},{"cp_plan":{"max_rank_tokens":33460,"summary_ms":0.00021979212760925293,"wave_count":0},"fallback_work":2250004992,"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":7257.98388671875,"n":8,"spread_pct":0.7572143580301622},{"cp_plan":{"max_rank_tokens":29364,"summary_ms":0.0002086162567138672,"wave_count":0},"fallback_work":1981274624,"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":6421.80419921875,"n":8,"spread_pct":0.7008897224003763},{"cp_plan":{"max_rank_tokens":36532,"summary_ms":0.00020489096641540527,"wave_count":0},"fallback_work":2451495424,"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":7875.4248046875,"n":8,"spread_pct":0.5142641734975854},{"cp_plan":{"max_rank_tokens":48820,"summary_ms":0.0004954636096954346,"wave_count":0},"fallback_work":3207813632,"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":10214.52978515625,"n":8,"spread_pct":0.4335601710893887}],"cell":"cal-hetero2|Qwen/Qwen3.5-27B|L64|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16756,"summary_ms":0.00027939677238464355,"wave_count":0},"fallback_work":1152498176,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":6164.97265625,"n":8,"spread_pct":0.20233862807329464},{"cp_plan":{"max_rank_tokens":17524,"summary_ms":0.0003520399332046509,"wave_count":0},"fallback_work":1202698752,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":6375.474365234375,"n":8,"spread_pct":0.8879930541320398},{"cp_plan":{"max_rank_tokens":25204,"summary_ms":0.0009704381227493286,"wave_count":0},"fallback_work":1656437248,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":8686.43603515625,"n":8,"spread_pct":0.3748548116363847}],"cell":"cal-hetero3|Qwen/Qwen3.5-27B|L64|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":9588,"summary_ms":5.118153989315033,"wave_count":1},"fallback_work":1152185408,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":3537.9012451171875,"n":8,"spread_pct":0.21778669304673542},{"cp_plan":{"max_rank_tokens":10356,"summary_ms":5.197932943701744,"wave_count":1},"fallback_work":1202283584,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":3698.09130859375,"n":8,"spread_pct":0.50316945859027},{"cp_plan":{"max_rank_tokens":14336,"summary_ms":6.393183022737503,"wave_count":1},"fallback_work":1655587904,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":4882.097900390625,"n":8,"spread_pct":0.43082305095350704}],"cell":"cal-hetero3|Qwen/Qwen3.5-27B|L64|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,2,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6144,"summary_ms":14.807816594839096,"wave_count":3},"fallback_work":1152766304,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":2186.87109375,"n":8,"spread_pct":0.4570063893369344},{"cp_plan":{"max_rank_tokens":6656,"summary_ms":13.914043083786964,"wave_count":3},"fallback_work":1202714976,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":2198.2589111328125,"n":8,"spread_pct":0.42336417142983757},{"cp_plan":{"max_rank_tokens":7168,"summary_ms":6.445828825235367,"wave_count":3},"fallback_work":1655703904,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":2851.797119140625,"n":8,"spread_pct":0.4107795582914796}],"cell":"cal-hetero3|Qwen/Qwen3.5-27B|L64|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":16756,"summary_ms":0.0003110617399215698,"wave_count":0},"fallback_work":1152498176,"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":3762.6983642578125,"n":8,"spread_pct":0.5403252819837773},{"cp_plan":{"max_rank_tokens":17524,"summary_ms":0.00038929283618927,"wave_count":0},"fallback_work":1202698752,"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":3839.73486328125,"n":8,"spread_pct":0.36478018133016743},{"cp_plan":{"max_rank_tokens":25204,"summary_ms":0.000521540641784668,"wave_count":0},"fallback_work":1656437248,"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":5165.781982421875,"n":8,"spread_pct":0.5813968051390603}],"cell":"cal-hetero3|Qwen/Qwen3.5-27B|L64|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19380,"summary_ms":0.00042654573917388916,"wave_count":0},"fallback_work":1325160960,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":7131.437255859375,"n":8,"spread_pct":0.2643992798858022},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.0003110617399215698,"wave_count":0},"fallback_work":1694521856,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":9072.0126953125,"n":8,"spread_pct":0.26369949388255276},{"cp_plan":{"max_rank_tokens":20916,"summary_ms":0.00040046870708465576,"wave_count":0},"fallback_work":1425791488,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":7623.623291015625,"n":8,"spread_pct":0.23385369816226748},{"cp_plan":{"max_rank_tokens":29108,"summary_ms":0.0010095536708831787,"wave_count":0},"fallback_work":1912920576,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":10163.119140625,"n":8,"spread_pct":0.4573349140590873}],"cell":"cal-hetero|Qwen/Qwen3.5-27B|L64|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":11264,"summary_ms":6.816098466515541,"wave_count":1},"fallback_work":1324820544,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":4147.7255859375,"n":8,"spread_pct":0.7081841493043415},{"cp_plan":{"max_rank_tokens":15872,"summary_ms":5.553767085075378,"wave_count":1},"fallback_work":1693780032,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":5109.38330078125,"n":8,"spread_pct":0.2900893605199609},{"cp_plan":{"max_rank_tokens":12800,"summary_ms":7.48346745967865,"wave_count":1},"fallback_work":1425311808,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":4363.8701171875,"n":8,"spread_pct":0.6150244270972383},{"cp_plan":{"max_rank_tokens":17408,"summary_ms":7.537985220551491,"wave_count":1},"fallback_work":1911916608,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":5531.674072265625,"n":8,"spread_pct":0.3365119826397491}],"cell":"cal-hetero|Qwen/Qwen3.5-27B|L64|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,2,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":6580,"summary_ms":6.076585501432419,"wave_count":3},"fallback_work":1325535072,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":2589.2896728515625,"n":8,"spread_pct":0.868888757827667},{"cp_plan":{"max_rank_tokens":9728,"summary_ms":6.522014737129211,"wave_count":2},"fallback_work":1694097248,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":2989.92578125,"n":8,"spread_pct":0.569408861267198},{"cp_plan":{"max_rank_tokens":7680,"summary_ms":6.129983812570572,"wave_count":1},"fallback_work":1425858400,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":2573.9095458984375,"n":8,"spread_pct":0.6554751668530746},{"cp_plan":{"max_rank_tokens":10752,"summary_ms":14.022095128893852,"wave_count":3},"fallback_work":1912004448,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":3332.44873046875,"n":8,"spread_pct":0.6515229993897307}],"cell":"cal-hetero|Qwen/Qwen3.5-27B|L64|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,4,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"cp_plan":{"max_rank_tokens":19380,"summary_ms":0.0005234032869338989,"wave_count":0},"fallback_work":1325160960,"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":4348.7333984375,"n":8,"spread_pct":0.725505338636901},{"cp_plan":{"max_rank_tokens":25012,"summary_ms":0.00029616057872772217,"wave_count":0},"fallback_work":1694521856,"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":5402.65625,"n":8,"spread_pct":0.7620672701506782},{"cp_plan":{"max_rank_tokens":20916,"summary_ms":0.00038929283618927,"wave_count":0},"fallback_work":1425791488,"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":4612.174560546875,"n":8,"spread_pct":0.5179057825419275},{"cp_plan":{"max_rank_tokens":29108,"summary_ms":0.0007078051567077637,"wave_count":0},"fallback_work":1912920576,"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":6005.18798828125,"n":8,"spread_pct":0.5478254682942199}],"cell":"cal-hetero|Qwen/Qwen3.5-27B|L64|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":48.0,"layers":64.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":17408,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":48,"hidden_size":5120,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":24,"num_query_groups":4},"hidden_size":5120,"model":"Qwen/Qwen3.5-27B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[2,1,1,1],"source":"82140580dcb0211d1a7c2b7d1b1b41cb2d2247e4118779c2ea0c6f67db9fedef","workload":{"kind":"cal-hetero","seed":7777}} + ], + "fit_arguments": { + "exclude_cells": ["cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g6","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g7","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g7","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g0","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g1","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g2","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g3","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g4","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g5","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g6","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g7","cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp1|cp1","cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-hetero2|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-hetero3|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-hetero|Qwen/Qwen3.5-27B|L64|tp2|cp2"], + "holdout": "", + "incumbent": "depth_one", + "objective": "regret", + "require_complete": 8, + "rerank_shapes": [], + "shortlist_size": 3, + "terms": ["token_per_rank","token_cp_exchange","token_tp_collective","gdn_token_per_rank","attention_token_cp_exchange","tiny_segment_per_layer","level_cp_per_layer","level_tp_per_layer","gdn_level","gdn_level_tp"] + }, + "integer_table_milli_us": {"attention_token_cp_exchange":0,"gdn_level":3043381,"gdn_level_tp":0,"gdn_token_per_rank":0,"level_cp_per_layer":0,"level_tp_per_layer":442039,"tiny_segment_per_layer":230182,"token_cp_exchange":0,"token_per_rank":11441,"token_tp_collective":0}, + "integer_table_sha256": "d01249d35845937d4cd21de199013286671efca1995def31aad197d49070bc38", + "manifest": { + "excluded": ["cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g6","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g7","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g7","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g0","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g1","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g2","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g3","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g4","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g5","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g6","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g7","cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp1|cp1","cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-hetero2|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-hetero3|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-hetero|Qwen/Qwen3.5-27B|L64|tp2|cp2"], + "expected": ["cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g0","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g1","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g2","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g3","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g4","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g5","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g6","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g7","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g0","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g1","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g2","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g3","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g4","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g5","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g6","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g7","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g0","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g1","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g2","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g3","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g4","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g5","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g6","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g7","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g0","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g1","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g2","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g3","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g4","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g5","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g6","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g7","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g0","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g1","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g2","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g3","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g4","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g5","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g6","cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g7","cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp1|cp1","cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp1|cp2","cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp1|cp4","cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp2|cp1","cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp1|cp1","cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp1|cp2","cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp1|cp4","cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp2|cp1","cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp1|cp1","cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp1|cp2","cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp1|cp4","cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp2|cp1","cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-hetero2|Qwen/Qwen3.5-27B|L64|tp1|cp1","cal-hetero2|Qwen/Qwen3.5-27B|L64|tp1|cp2","cal-hetero2|Qwen/Qwen3.5-27B|L64|tp1|cp4","cal-hetero2|Qwen/Qwen3.5-27B|L64|tp2|cp1","cal-hetero2|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-hetero3|Qwen/Qwen3.5-27B|L64|tp1|cp1","cal-hetero3|Qwen/Qwen3.5-27B|L64|tp1|cp2","cal-hetero3|Qwen/Qwen3.5-27B|L64|tp1|cp4","cal-hetero3|Qwen/Qwen3.5-27B|L64|tp2|cp1","cal-hetero3|Qwen/Qwen3.5-27B|L64|tp2|cp2","cal-hetero|Qwen/Qwen3.5-27B|L64|tp1|cp1","cal-hetero|Qwen/Qwen3.5-27B|L64|tp1|cp2","cal-hetero|Qwen/Qwen3.5-27B|L64|tp1|cp4","cal-hetero|Qwen/Qwen3.5-27B|L64|tp2|cp1","cal-hetero|Qwen/Qwen3.5-27B|L64|tp2|cp2"], + "path": "trainer_rank_cost_calibration_manifest_dense-gdn-h5120-h200-bf16.json" + }, + "measured_envelope": {"device_names":["NVIDIA H200"],"hidden_sizes":[5120],"param_dtypes":["torch.bfloat16"]}, + "metrics": { + "integer_all": {"cells":52,"clear_misses":[],"max_regret_pct":3.3957125091334266,"median_regret_pct":0.0,"ordered_pairs":3464,"p95_regret_pct":0.8748881696713853,"pairwise_accuracy":0.9699769053117783}, + "test": {"cells":14,"clear_misses":[],"max_regret_pct":1.810408206308734,"median_regret_pct":0.0,"ordered_pairs":613,"p95_regret_pct":1.810408206308734,"pairwise_accuracy":1.0}, + "train": {"cells":38,"clear_misses":[],"max_regret_pct":3.3957125091334266,"median_regret_pct":0.0,"ordered_pairs":2851,"p95_regret_pct":0.8748881696713853,"pairwise_accuracy":0.9635215713784637} + }, + "reranker": null, + "schema": "art.dev.trainer_rank_cost_calibration_certificate.v1", + "table_id": "dense-gdn-h5120-h200-bf16" +} diff --git a/dev/trainer_rank_cost_calibration_certificate.json b/dev/trainer_rank_cost_calibration_certificate_dense-h2560-h200-bf16.json similarity index 98% rename from dev/trainer_rank_cost_calibration_certificate.json rename to dev/trainer_rank_cost_calibration_certificate_dense-h2560-h200-bf16.json index f59c78478..1cb582bf2 100644 --- a/dev/trainer_rank_cost_calibration_certificate.json +++ b/dev/trainer_rank_cost_calibration_certificate_dense-h2560-h200-bf16.json @@ -61,7 +61,7 @@ {"candidates":[{"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1107.144287109375,"n":8,"spread_pct":4.335775965007102},{"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1305.4077758789062,"n":8,"spread_pct":2.64851912545848},{"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1173.864013671875,"n":8,"spread_pct":3.677009043192731},{"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1321.0035400390625,"n":8,"spread_pct":1.2978506719070204}],"cell":"cal-hetero|Qwen/Qwen3.5-4B|L32|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":24.0,"layers":32.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":9216,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2560,"kv_channels":256,"moe_experts":0,"moe_ffn_hidden_size":0,"moe_shared_expert_ffn":0,"moe_topk":0,"num_attention_heads":16,"num_query_groups":4},"hidden_size":2560,"model":"Qwen/Qwen3.5-4B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[2,1,1,1],"source":"797746b9f425d844419fe609c164c2e4940507f3bbb2f82b815a73740d5abc2d","workload":{"kind":"hetero","seed":7777}} ], "fit_arguments": { - "exclude_cells": [], + "exclude_cells": ["cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g0","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g1","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g2","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g3","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g4","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g5","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g6","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g7","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g7"], "holdout": "cp4|g4", "objective": "regret", "require_complete": 8, @@ -70,9 +70,9 @@ "integer_table_milli_us": {"attention_token_cp_exchange":27,"gdn_level":3336555,"gdn_level_tp":1194872,"gdn_token_per_rank":594,"level_cp_per_layer":187255,"level_tp_per_layer":684798,"tiny_segment_per_layer":219288,"token_cp_exchange":39,"token_per_rank":2002,"token_tp_collective":221}, "integer_table_sha256": "3624ae7dd985d92fed47b511e058f0e792559ead8919e11585dcb9860f3394ad", "manifest": { - "excluded": [], - "expected": ["cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g0","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g1","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g2","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g3","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g4","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g5","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g6","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g7","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g0","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g1","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g2","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g3","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g4","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g5","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g6","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g7","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g0","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g1","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g2","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g3","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g4","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g5","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g6","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g7","cal-grpo-g16|Qwen/Qwen3.5-4B|L32|tp1|cp1","cal-grpo-g16|Qwen/Qwen3.5-4B|L32|tp1|cp2","cal-grpo-g16|Qwen/Qwen3.5-4B|L32|tp1|cp4","cal-grpo-g16|Qwen/Qwen3.5-4B|L32|tp2|cp1","cal-grpo-g4x4|Qwen/Qwen3.5-4B|L32|tp1|cp1","cal-grpo-g4x4|Qwen/Qwen3.5-4B|L32|tp1|cp2","cal-grpo-g4x4|Qwen/Qwen3.5-4B|L32|tp1|cp4","cal-grpo-g4x4|Qwen/Qwen3.5-4B|L32|tp2|cp1","cal-grpo-g8-long|Qwen/Qwen3-4B|L2|tp1|cp1","cal-grpo-g8-long|Qwen/Qwen3-4B|L2|tp1|cp4","cal-grpo-g8-long|Qwen/Qwen3.5-4B|L2|tp1|cp1","cal-grpo-g8-long|Qwen/Qwen3.5-4B|L2|tp1|cp4","cal-grpo-g8|Qwen/Qwen3-4B|L36|tp1|cp1","cal-grpo-g8|Qwen/Qwen3-4B|L36|tp1|cp2","cal-grpo-g8|Qwen/Qwen3-4B|L36|tp1|cp4","cal-grpo-g8|Qwen/Qwen3-4B|L36|tp2|cp1","cal-grpo-g8|Qwen/Qwen3.5-4B|L32|tp1|cp1","cal-grpo-g8|Qwen/Qwen3.5-4B|L32|tp1|cp2","cal-grpo-g8|Qwen/Qwen3.5-4B|L32|tp1|cp4","cal-grpo-g8|Qwen/Qwen3.5-4B|L32|tp2|cp1","cal-hetero2|Qwen/Qwen3-4B|L36|tp1|cp1","cal-hetero2|Qwen/Qwen3-4B|L36|tp1|cp4","cal-hetero2|Qwen/Qwen3.5-4B|L32|tp1|cp1","cal-hetero2|Qwen/Qwen3.5-4B|L32|tp1|cp2","cal-hetero2|Qwen/Qwen3.5-4B|L32|tp1|cp4","cal-hetero2|Qwen/Qwen3.5-4B|L32|tp2|cp1","cal-hetero3|Qwen/Qwen3.5-4B|L32|tp1|cp1","cal-hetero3|Qwen/Qwen3.5-4B|L32|tp1|cp2","cal-hetero3|Qwen/Qwen3.5-4B|L32|tp1|cp4","cal-hetero3|Qwen/Qwen3.5-4B|L32|tp2|cp1","cal-hetero|Qwen/Qwen3.5-4B|L32|tp1|cp1","cal-hetero|Qwen/Qwen3.5-4B|L32|tp1|cp2","cal-hetero|Qwen/Qwen3.5-4B|L32|tp1|cp4","cal-hetero|Qwen/Qwen3.5-4B|L32|tp2|cp1"], - "path": "trainer_rank_cost_calibration_manifest.json" + "excluded": ["cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g0","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g1","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g2","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g3","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g4","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g5","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g6","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g7","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g7"], + "expected": ["cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g0","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g1","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g2","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g3","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g4","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g5","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g6","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g7","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g0","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g1","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g2","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g3","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g4","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g5","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g6","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp2|g7","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g0","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g1","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g2","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g3","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g4","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g5","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g6","cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp4|g7","cal-grpo-g16|Qwen/Qwen3.5-4B|L32|tp1|cp1","cal-grpo-g16|Qwen/Qwen3.5-4B|L32|tp1|cp2","cal-grpo-g16|Qwen/Qwen3.5-4B|L32|tp1|cp4","cal-grpo-g16|Qwen/Qwen3.5-4B|L32|tp2|cp1","cal-grpo-g4x4|Qwen/Qwen3.5-4B|L32|tp1|cp1","cal-grpo-g4x4|Qwen/Qwen3.5-4B|L32|tp1|cp2","cal-grpo-g4x4|Qwen/Qwen3.5-4B|L32|tp1|cp4","cal-grpo-g4x4|Qwen/Qwen3.5-4B|L32|tp2|cp1","cal-grpo-g8-long|Qwen/Qwen3-4B|L2|tp1|cp1","cal-grpo-g8-long|Qwen/Qwen3-4B|L2|tp1|cp4","cal-grpo-g8-long|Qwen/Qwen3.5-4B|L2|tp1|cp1","cal-grpo-g8-long|Qwen/Qwen3.5-4B|L2|tp1|cp4","cal-grpo-g8|Qwen/Qwen3-4B|L36|tp1|cp1","cal-grpo-g8|Qwen/Qwen3-4B|L36|tp1|cp2","cal-grpo-g8|Qwen/Qwen3-4B|L36|tp1|cp4","cal-grpo-g8|Qwen/Qwen3-4B|L36|tp2|cp1","cal-grpo-g8|Qwen/Qwen3.5-4B|L32|tp1|cp1","cal-grpo-g8|Qwen/Qwen3.5-4B|L32|tp1|cp2","cal-grpo-g8|Qwen/Qwen3.5-4B|L32|tp1|cp4","cal-grpo-g8|Qwen/Qwen3.5-4B|L32|tp2|cp1","cal-hetero2|Qwen/Qwen3-4B|L36|tp1|cp1","cal-hetero2|Qwen/Qwen3-4B|L36|tp1|cp4","cal-hetero2|Qwen/Qwen3.5-4B|L32|tp1|cp1","cal-hetero2|Qwen/Qwen3.5-4B|L32|tp1|cp2","cal-hetero2|Qwen/Qwen3.5-4B|L32|tp1|cp4","cal-hetero2|Qwen/Qwen3.5-4B|L32|tp2|cp1","cal-hetero3|Qwen/Qwen3.5-4B|L32|tp1|cp1","cal-hetero3|Qwen/Qwen3.5-4B|L32|tp1|cp2","cal-hetero3|Qwen/Qwen3.5-4B|L32|tp1|cp4","cal-hetero3|Qwen/Qwen3.5-4B|L32|tp2|cp1","cal-hetero|Qwen/Qwen3.5-4B|L32|tp1|cp1","cal-hetero|Qwen/Qwen3.5-4B|L32|tp1|cp2","cal-hetero|Qwen/Qwen3.5-4B|L32|tp1|cp4","cal-hetero|Qwen/Qwen3.5-4B|L32|tp2|cp1","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g0","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g1","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g2","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g3","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g4","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g5","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g6","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g7","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g0","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g1","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g2","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g3","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g4","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g5","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g6","cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g7"], + "path": "trainer_rank_cost_calibration_manifest_dense-h2560-h200-bf16.json" }, "measured_envelope": {"device_names":["NVIDIA H200"],"hidden_sizes":[2560],"param_dtypes":["torch.bfloat16"]}, "metrics": { diff --git a/dev/trainer_rank_cost_calibration_certificate_gdn-moe-h2048-h200-bf16.json b/dev/trainer_rank_cost_calibration_certificate_gdn-moe-h2048-h200-bf16.json new file mode 100644 index 000000000..ff084d24f --- /dev/null +++ b/dev/trainer_rank_cost_calibration_certificate_gdn-moe-h2048-h200-bf16.json @@ -0,0 +1,100 @@ +{ + "admitted": {"device_classes":[{"capability":[9,0],"memory_class":"hbm-141g"}],"geometries":[{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2}],"param_dtypes":["torch.bfloat16"],"shapes":[[1,1,1,1],[1,2,1,1],[1,2,2,1],[1,4,1,1],[1,4,2,1],[1,4,4,1],[2,1,1,1],[2,1,2,1]]}, + "cells": [ + {"candidates":[{"features":{"max_depth":2,"packed_tokens":12659,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1730.8432006835938,"n":8,"spread_pct":6.146828767124495},{"features":{"max_depth":2,"packed_tokens":24836,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12268","median_ms":2760.64404296875,"n":8,"spread_pct":4.962309413546538},{"features":{"max_depth":2,"packed_tokens":37143,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12313","median_ms":3916.5966796875,"n":8,"spread_pct":21.55801695259681},{"features":{"max_depth":3,"packed_tokens":12609,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1932.1292724609375,"n":8,"spread_pct":5.722136898761012},{"features":{"max_depth":1,"packed_tokens":49460,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":4739.919189453125,"n":8,"spread_pct":0.7595066194399308},{"features":{"max_depth":3,"packed_tokens":12569,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1942.1019897460938,"n":8,"spread_pct":5.091925212038103},{"features":{"max_depth":4,"packed_tokens":12564,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":2070.2525634765625,"n":8,"spread_pct":6.133632760240877}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"f8b37d8f36ed38a4d2ea74f20bfe0d15bedbdb66ffaae41eb728a1756d67d5ff","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":0,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":12382,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1607.5150756835938,"n":8,"spread_pct":1.1657510593154778},{"features":{"max_depth":1,"packed_tokens":24645,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2452.4344482421875,"n":8,"spread_pct":29.272441988670877}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"515df421b93fe5f98e250ae3d849975dccadfd7dc00046242b5875ddf392c74f","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":1,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":1,"packed_tokens":12546,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1531.3175659179688,"n":8,"spread_pct":1.1008442827399159}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8aa85999c5d3c576d127bd2e3a44436cbd0e9062c7c92b16bcbaab0814703094","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":2,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":47253,"segment_count":6,"segments_below":[0,0,0,1,1,1,1,3]},"label":"depth_one","median_ms":4685.706787109375,"n":8,"spread_pct":0.24759471787514548},{"features":{"max_depth":2,"packed_tokens":23150,"segment_count":7,"segments_below":[0,0,0,1,3,5,5,6]},"label":"minimum_effective_span_1366","median_ms":2480.6708984375,"n":8,"spread_pct":0.8845046880037729},{"features":{"max_depth":3,"packed_tokens":21785,"segment_count":7,"segments_below":[0,0,0,1,4,5,5,6]},"label":"minimum_effective_span_305","median_ms":2458.5634765625,"n":8,"spread_pct":1.011034851477755},{"features":{"max_depth":2,"packed_tokens":28799,"segment_count":6,"segments_below":[0,0,0,1,3,3,3,4]},"label":"minimum_effective_span_7015","median_ms":3041.6375732421875,"n":8,"spread_pct":0.6538643449406923},{"features":{"max_depth":2,"packed_tokens":37178,"segment_count":6,"segments_below":[0,0,0,1,2,2,2,3]},"label":"minimum_effective_span_8380","median_ms":3810.885498046875,"n":8,"spread_pct":0.1887710260499019},{"features":{"max_depth":1,"packed_tokens":48469,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":4570.166259765625,"n":8,"spread_pct":0.1607102705903891},{"features":{"max_depth":3,"packed_tokens":22846,"segment_count":8,"segments_below":[0,0,0,2,4,6,6,7]},"label":"uniform_depth_2","median_ms":2526.79150390625,"n":8,"spread_pct":28.510422196194725},{"features":{"max_depth":4,"packed_tokens":21481,"segment_count":8,"segments_below":[0,0,0,2,5,6,6,7]},"label":"uniform_depth_3","median_ms":2505.1583251953125,"n":8,"spread_pct":0.7759770282576619}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"62112dd684879b09adbef34aa98e5ea3966b7de9b9e8067de5920b1d73f07197","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":3,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":30477,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,2]},"label":"depth_one","median_ms":3118.3458251953125,"n":8,"spread_pct":0.37917457755457334},{"features":{"max_depth":2,"packed_tokens":19821,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_495","median_ms":2218.517578125,"n":8,"spread_pct":1.2238738036819},{"features":{"max_depth":1,"packed_tokens":31465,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":3041.143310546875,"n":8,"spread_pct":0.4278240404645823},{"features":{"max_depth":3,"packed_tokens":19327,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,4]},"label":"uniform_depth_2","median_ms":2228.2384033203125,"n":8,"spread_pct":0.9759101825166792}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"28b4492a5c1ce50958108e27b72b7afdf09941e4670076dfa879b517d9cc7575","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":4,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":12659,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1524.919677734375,"n":8,"spread_pct":4.631831754714284},{"features":{"max_depth":2,"packed_tokens":24836,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12268","median_ms":1731.2490234375,"n":8,"spread_pct":3.731824645659209},{"features":{"max_depth":2,"packed_tokens":37143,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12313","median_ms":2556.45068359375,"n":8,"spread_pct":1.9790908835664072},{"features":{"max_depth":3,"packed_tokens":12609,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1797.651123046875,"n":8,"spread_pct":50.09846290576394},{"features":{"max_depth":1,"packed_tokens":49460,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2779.2601318359375,"n":8,"spread_pct":0.9214633415524335},{"features":{"max_depth":3,"packed_tokens":12569,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1797.0164184570312,"n":8,"spread_pct":7.530847521273547},{"features":{"max_depth":4,"packed_tokens":12564,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1929.5479125976562,"n":8,"spread_pct":7.828817829781652}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"f8b37d8f36ed38a4d2ea74f20bfe0d15bedbdb66ffaae41eb728a1756d67d5ff","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":0,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":12382,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1345.421875,"n":8,"spread_pct":1.7101088616488787},{"features":{"max_depth":1,"packed_tokens":24645,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1554.2542114257812,"n":8,"spread_pct":0.5343276221651126}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"515df421b93fe5f98e250ae3d849975dccadfd7dc00046242b5875ddf392c74f","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":1,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":1,"packed_tokens":12546,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1308.4957885742188,"n":8,"spread_pct":3.805947226926269}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8aa85999c5d3c576d127bd2e3a44436cbd0e9062c7c92b16bcbaab0814703094","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":2,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":47253,"segment_count":6,"segments_below":[0,0,0,1,1,1,1,3]},"label":"depth_one","median_ms":2762.767578125,"n":8,"spread_pct":25.863135782238288},{"features":{"max_depth":2,"packed_tokens":23150,"segment_count":7,"segments_below":[0,0,0,1,3,5,5,6]},"label":"minimum_effective_span_1366","median_ms":1665.4863891601562,"n":8,"spread_pct":45.15628480896756},{"features":{"max_depth":3,"packed_tokens":21785,"segment_count":7,"segments_below":[0,0,0,1,4,5,5,6]},"label":"minimum_effective_span_305","median_ms":1699.326416015625,"n":8,"spread_pct":2.254584156377448},{"features":{"max_depth":2,"packed_tokens":28799,"segment_count":6,"segments_below":[0,0,0,1,3,3,3,4]},"label":"minimum_effective_span_7015","median_ms":1782.0517578125,"n":8,"spread_pct":0.9314954727066381},{"features":{"max_depth":2,"packed_tokens":37178,"segment_count":6,"segments_below":[0,0,0,1,2,2,2,3]},"label":"minimum_effective_span_8380","median_ms":2237.9251708984375,"n":8,"spread_pct":33.2296100988579},{"features":{"max_depth":1,"packed_tokens":48469,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":2787.8466796875,"n":8,"spread_pct":0.8203943946152998},{"features":{"max_depth":3,"packed_tokens":22846,"segment_count":8,"segments_below":[0,0,0,2,4,6,6,7]},"label":"uniform_depth_2","median_ms":1708.3840942382812,"n":8,"spread_pct":0.99765042471168},{"features":{"max_depth":4,"packed_tokens":21481,"segment_count":8,"segments_below":[0,0,0,2,5,6,6,7]},"label":"uniform_depth_3","median_ms":1742.3125,"n":8,"spread_pct":2.0468752802489507}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"62112dd684879b09adbef34aa98e5ea3966b7de9b9e8067de5920b1d73f07197","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":3,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":30477,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,2]},"label":"depth_one","median_ms":1918.074462890625,"n":8,"spread_pct":1.250828461211341},{"features":{"max_depth":2,"packed_tokens":19821,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_495","median_ms":1448.9666748046875,"n":8,"spread_pct":49.21205549988269},{"features":{"max_depth":1,"packed_tokens":31465,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":1949.8007202148438,"n":8,"spread_pct":1.7883250350948887},{"features":{"max_depth":3,"packed_tokens":19327,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,4]},"label":"uniform_depth_2","median_ms":1483.6954345703125,"n":8,"spread_pct":49.35803257955349}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"28b4492a5c1ce50958108e27b72b7afdf09941e4670076dfa879b517d9cc7575","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":4,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":13557,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1421.724609375,"n":8,"spread_pct":3.2049148742380718},{"features":{"max_depth":2,"packed_tokens":25902,"segment_count":6,"segments_below":[0,1,2,4,4,4,4,4]},"label":"minimum_effective_span_12406","median_ms":1684.3063354492188,"n":8,"spread_pct":2.11820874104366},{"features":{"max_depth":2,"packed_tokens":37995,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12426","median_ms":2530.314697265625,"n":8,"spread_pct":1.7263226796089488},{"features":{"max_depth":2,"packed_tokens":50582,"segment_count":6,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12592","median_ms":2886.303466796875,"n":8,"spread_pct":0.6733969425066986},{"features":{"max_depth":3,"packed_tokens":13367,"segment_count":7,"segments_below":[1,3,5,6,6,6,6,6]},"label":"minimum_effective_span_187","median_ms":1669.5917358398438,"n":8,"spread_pct":173.8413343485925},{"features":{"max_depth":3,"packed_tokens":13185,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_21","median_ms":1669.8106689453125,"n":8,"spread_pct":3.616439662027201},{"features":{"max_depth":1,"packed_tokens":63177,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3656.0050048828125,"n":8,"spread_pct":1.0042807021691125},{"features":{"max_depth":3,"packed_tokens":13497,"segment_count":7,"segments_below":[1,2,4,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1684.7539672851562,"n":8,"spread_pct":44.55749733661283},{"features":{"max_depth":4,"packed_tokens":13165,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1777.91064453125,"n":8,"spread_pct":4.275785937507296},{"features":{"max_depth":5,"packed_tokens":13161,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1918.0214233398438,"n":8,"spread_pct":2.489773781989121}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"7241823ac1b67d127c5921d771dda170dcc9d4886feffaff3c2520113cd01a08","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":5,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":9433,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1171.02294921875,"n":8,"spread_pct":64.6748886169319},{"features":{"max_depth":6,"packed_tokens":6738,"segment_count":17,"segments_below":[6,11,15,16,16,16,16,17]},"label":"minimum_effective_span_105","median_ms":1723.1726684570312,"n":8,"spread_pct":44.26256003683136},{"features":{"max_depth":6,"packed_tokens":6778,"segment_count":17,"segments_below":[5,10,15,16,16,16,16,17]},"label":"minimum_effective_span_113","median_ms":1734.2799072265625,"n":8,"spread_pct":221.38104278099144},{"features":{"max_depth":6,"packed_tokens":6877,"segment_count":17,"segments_below":[5,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":1739.51171875,"n":8,"spread_pct":5.2578231139755},{"features":{"max_depth":5,"packed_tokens":7090,"segment_count":16,"segments_below":[3,8,14,15,15,15,15,16]},"label":"minimum_effective_span_164","median_ms":1602.9992065429688,"n":8,"spread_pct":48.60788182430403},{"features":{"max_depth":5,"packed_tokens":7134,"segment_count":16,"segments_below":[4,7,14,15,15,15,15,16]},"label":"minimum_effective_span_195","median_ms":1617.677490234375,"n":8,"spread_pct":7.5115752216983775},{"features":{"max_depth":4,"packed_tokens":7415,"segment_count":16,"segments_below":[3,5,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":1498.873779296875,"n":8,"spread_pct":4.328151979994094},{"features":{"max_depth":4,"packed_tokens":7626,"segment_count":15,"segments_below":[3,6,10,14,14,14,14,15]},"label":"minimum_effective_span_245","median_ms":1502.9976806640625,"n":8,"spread_pct":5.815412671795143},{"features":{"max_depth":4,"packed_tokens":7702,"segment_count":15,"segments_below":[3,5,9,14,14,14,14,15]},"label":"minimum_effective_span_281","median_ms":1483.4010620117188,"n":8,"spread_pct":50.91368543085609},{"features":{"max_depth":12,"packed_tokens":6323,"segment_count":23,"segments_below":[16,20,22,22,22,22,22,23]},"label":"minimum_effective_span_32","median_ms":2438.5057373046875,"n":8,"spread_pct":5.202701069005031},{"features":{"max_depth":3,"packed_tokens":8041,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_340","median_ms":1414.3967895507812,"n":8,"spread_pct":51.8396181979942},{"features":{"max_depth":3,"packed_tokens":8113,"segment_count":15,"segments_below":[1,4,8,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":1368.5878295898438,"n":8,"spread_pct":4.233565593954371},{"features":{"max_depth":11,"packed_tokens":6359,"segment_count":22,"segments_below":[15,19,21,21,21,21,21,22]},"label":"minimum_effective_span_37","median_ms":2328.0228271484375,"n":8,"spread_pct":5.374889827117517},{"features":{"max_depth":9,"packed_tokens":6437,"segment_count":20,"segments_below":[9,17,19,19,19,19,19,20]},"label":"minimum_effective_span_40","median_ms":2094.4752197265625,"n":8,"spread_pct":3.485126799305348},{"features":{"max_depth":3,"packed_tokens":8391,"segment_count":15,"segments_below":[1,3,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":1361.9873657226562,"n":8,"spread_pct":68.9357435900952},{"features":{"max_depth":2,"packed_tokens":13958,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4955","median_ms":1304.1620483398438,"n":8,"spread_pct":4.791057551200331},{"features":{"max_depth":2,"packed_tokens":18411,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4994","median_ms":1546.286376953125,"n":8,"spread_pct":3.5862131249754285},{"features":{"max_depth":2,"packed_tokens":23107,"segment_count":14,"segments_below":[2,3,5,7,10,10,10,14]},"label":"minimum_effective_span_5048","median_ms":1590.351806640625,"n":8,"spread_pct":46.79169632677272},{"features":{"max_depth":2,"packed_tokens":27945,"segment_count":14,"segments_below":[2,3,5,6,9,9,9,14]},"label":"minimum_effective_span_5087","median_ms":1882.5144653320312,"n":8,"spread_pct":2.126775967552177},{"features":{"max_depth":2,"packed_tokens":32845,"segment_count":14,"segments_below":[1,3,4,6,8,8,8,14]},"label":"minimum_effective_span_5118","median_ms":1975.6542358398438,"n":8,"spread_pct":1.5828481460987556},{"features":{"max_depth":2,"packed_tokens":37693,"segment_count":14,"segments_below":[2,2,3,6,7,7,7,14]},"label":"minimum_effective_span_5149","median_ms":2259.0634765625,"n":8,"spread_pct":1.1720474079889127},{"features":{"max_depth":2,"packed_tokens":42711,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_5199","median_ms":2356.822998046875,"n":8,"spread_pct":1.7403965609739094},{"features":{"max_depth":3,"packed_tokens":8746,"segment_count":14,"segments_below":[2,4,7,10,13,13,13,14]},"label":"minimum_effective_span_522","median_ms":1241.2423095703125,"n":8,"spread_pct":338.6459744280657},{"features":{"max_depth":2,"packed_tokens":47673,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5235","median_ms":2703.5897216796875,"n":8,"spread_pct":1.7956269891642094},{"features":{"max_depth":2,"packed_tokens":52699,"segment_count":14,"segments_below":[0,1,2,4,4,4,4,14]},"label":"minimum_effective_span_5303","median_ms":2805.9219970703125,"n":8,"spread_pct":23.44297185000352},{"features":{"max_depth":2,"packed_tokens":57931,"segment_count":14,"segments_below":[0,0,3,3,3,3,3,14]},"label":"minimum_effective_span_5395","median_ms":3147.7470703125,"n":8,"spread_pct":1.5481450181537948},{"features":{"max_depth":2,"packed_tokens":63240,"segment_count":13,"segments_below":[1,1,1,1,1,1,1,13]},"label":"minimum_effective_span_5476","median_ms":3212.7955322265625,"n":8,"spread_pct":21.683977652682046},{"features":{"max_depth":8,"packed_tokens":6558,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_71","median_ms":1959.0301513671875,"n":8,"spread_pct":40.53953749211369},{"features":{"max_depth":7,"packed_tokens":6639,"segment_count":18,"segments_below":[6,14,16,17,17,17,17,18]},"label":"minimum_effective_span_82","median_ms":1890.7329711914062,"n":8,"spread_pct":9.395344125653464},{"features":{"max_depth":7,"packed_tokens":6657,"segment_count":18,"segments_below":[6,14,16,17,17,17,17,18]},"label":"minimum_effective_span_87","median_ms":1859.6332397460938,"n":8,"spread_pct":6.526076292384681},{"features":{"max_depth":7,"packed_tokens":6668,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_93","median_ms":1857.3670043945312,"n":8,"spread_pct":40.85389483607577},{"features":{"max_depth":7,"packed_tokens":6645,"segment_count":18,"segments_below":[7,12,17,17,17,17,17,18]},"label":"minimum_effective_span_94","median_ms":1829.8212890625,"n":8,"spread_pct":4.708725130581293},{"features":{"max_depth":1,"packed_tokens":68881,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":3575.2843017578125,"n":8,"spread_pct":0.32402908299639765},{"features":{"max_depth":11,"packed_tokens":6620,"segment_count":23,"segments_below":[16,19,20,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":2331.7060546875,"n":8,"spread_pct":32.99583861568636},{"features":{"max_depth":12,"packed_tokens":6458,"segment_count":24,"segments_below":[16,20,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":2478.818603515625,"n":8,"spread_pct":3.5346545994525296},{"features":{"max_depth":13,"packed_tokens":6292,"segment_count":24,"segments_below":[17,21,23,23,23,23,23,24]},"label":"uniform_depth_12","median_ms":2572.2779541015625,"n":8,"spread_pct":3.845814899995022},{"features":{"max_depth":3,"packed_tokens":9004,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1371.4413452148438,"n":8,"spread_pct":7.4245511565688815},{"features":{"max_depth":4,"packed_tokens":8464,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1493.2850952148438,"n":8,"spread_pct":4.830911555852875},{"features":{"max_depth":5,"packed_tokens":8113,"segment_count":17,"segments_below":[8,9,11,13,16,16,16,17]},"label":"uniform_depth_4","median_ms":1626.0341186523438,"n":8,"spread_pct":44.92074473489921},{"features":{"max_depth":6,"packed_tokens":7865,"segment_count":18,"segments_below":[10,11,13,14,17,17,17,18]},"label":"uniform_depth_5","median_ms":1752.3919677734375,"n":8,"spread_pct":8.376113982039755},{"features":{"max_depth":7,"packed_tokens":7648,"segment_count":19,"segments_below":[11,13,14,16,18,18,18,19]},"label":"uniform_depth_6","median_ms":1897.1909790039062,"n":8,"spread_pct":63.0522799869526},{"features":{"max_depth":8,"packed_tokens":7348,"segment_count":20,"segments_below":[14,14,15,18,19,19,19,20]},"label":"uniform_depth_7","median_ms":2003.7261962890625,"n":8,"spread_pct":185.85655082921986},{"features":{"max_depth":9,"packed_tokens":7168,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":2118.3389892578125,"n":8,"spread_pct":34.163718498812315},{"features":{"max_depth":10,"packed_tokens":6896,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":2218.544189453125,"n":8,"spread_pct":4.458721248020696}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"22fe33a263cacb6bc86ab088fe07af0a0fb8e68bfdc85c518a57ec37e403d138","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":6,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":12674,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1444.2356567382812,"n":8,"spread_pct":8.166506827701062},{"features":{"max_depth":2,"packed_tokens":24329,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11722","median_ms":1634.4521484375,"n":8,"spread_pct":60.169206580234324},{"features":{"max_depth":2,"packed_tokens":36051,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11733","median_ms":2481.9210205078125,"n":8,"spread_pct":2.603347833210054},{"features":{"max_depth":2,"packed_tokens":47693,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11735","median_ms":2765.8946533203125,"n":8,"spread_pct":1.7128714807070584},{"features":{"max_depth":2,"packed_tokens":59300,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11758","median_ms":3573.02685546875,"n":8,"spread_pct":30.58035473304691},{"features":{"max_depth":2,"packed_tokens":71013,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11808","median_ms":3920.2518310546875,"n":8,"spread_pct":0.9202451599817068},{"features":{"max_depth":2,"packed_tokens":82862,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11855","median_ms":4606.249267578125,"n":8,"spread_pct":0.5512636561753185},{"features":{"max_depth":6,"packed_tokens":12273,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":2124.8822021484375,"n":8,"spread_pct":50.24468555385797},{"features":{"max_depth":3,"packed_tokens":12536,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1717.37841796875,"n":8,"spread_pct":5.13510639598475},{"features":{"max_depth":5,"packed_tokens":12286,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1995.337158203125,"n":8,"spread_pct":38.81370886543224},{"features":{"max_depth":7,"packed_tokens":12259,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":2232.7498779296875,"n":8,"spread_pct":4.480979856032745},{"features":{"max_depth":4,"packed_tokens":12322,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1854.150146484375,"n":8,"spread_pct":40.57987788427012},{"features":{"max_depth":4,"packed_tokens":12364,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1840.532958984375,"n":8,"spread_pct":12.461089748359923},{"features":{"max_depth":3,"packed_tokens":12416,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1719.4125366210938,"n":8,"spread_pct":6.7998859104560365},{"features":{"max_depth":6,"packed_tokens":12264,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":2101.676025390625,"n":8,"spread_pct":9.277043003601687},{"features":{"max_depth":3,"packed_tokens":12408,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1730.2884521484375,"n":8,"spread_pct":6.118463386739716},{"features":{"max_depth":1,"packed_tokens":94721,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":5003.524169921875,"n":8,"spread_pct":0.6544215539486695},{"features":{"max_depth":3,"packed_tokens":12608,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1703.0924072265625,"n":8,"spread_pct":68.95562723283194},{"features":{"max_depth":4,"packed_tokens":12598,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1873.6323852539062,"n":8,"spread_pct":5.928314965932345},{"features":{"max_depth":5,"packed_tokens":12506,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1988.1563720703125,"n":8,"spread_pct":37.67439470197569},{"features":{"max_depth":6,"packed_tokens":12356,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":2128.9368896484375,"n":8,"spread_pct":77.82501524777366},{"features":{"max_depth":7,"packed_tokens":12262,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":2257.2506103515625,"n":8,"spread_pct":7.980511153375811},{"features":{"max_depth":8,"packed_tokens":12257,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":2372.608642578125,"n":8,"spread_pct":7.885437691425526}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"d04b77db2ed7d695d20c6f183215ead28d0337ae88b9ddbc3674926310d364b9","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":7,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":12659,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1555.8897705078125,"n":8,"spread_pct":2.0368278331243372},{"features":{"max_depth":2,"packed_tokens":24836,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12268","median_ms":1745.2158813476562,"n":8,"spread_pct":4.624332766295359},{"features":{"max_depth":2,"packed_tokens":37143,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12313","median_ms":2545.4935302734375,"n":8,"spread_pct":31.15878182198425},{"features":{"max_depth":3,"packed_tokens":12609,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1718.0316162109375,"n":8,"spread_pct":3.6053421476234915},{"features":{"max_depth":1,"packed_tokens":49460,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2674.8875732421875,"n":8,"spread_pct":0.8851587704086753},{"features":{"max_depth":3,"packed_tokens":12569,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1713.55712890625,"n":8,"spread_pct":2.0479488454463985},{"features":{"max_depth":4,"packed_tokens":12564,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1835.1715698242188,"n":8,"spread_pct":42.36786816414867}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"f8b37d8f36ed38a4d2ea74f20bfe0d15bedbdb66ffaae41eb728a1756d67d5ff","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":0,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":12382,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1541.3357543945312,"n":8,"spread_pct":49.94193815487871},{"features":{"max_depth":1,"packed_tokens":24645,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1673.3619384765625,"n":8,"spread_pct":2.0255495605142655}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"515df421b93fe5f98e250ae3d849975dccadfd7dc00046242b5875ddf392c74f","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":1,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":47253,"segment_count":6,"segments_below":[0,0,0,1,1,1,1,3]},"label":"depth_one","median_ms":2807.2630615234375,"n":8,"spread_pct":3.454809669689207},{"features":{"max_depth":2,"packed_tokens":23150,"segment_count":7,"segments_below":[0,0,0,1,3,5,5,6]},"label":"minimum_effective_span_1366","median_ms":1849.7772827148438,"n":8,"spread_pct":37.97207427480432},{"features":{"max_depth":3,"packed_tokens":21785,"segment_count":7,"segments_below":[0,0,0,1,4,5,5,6]},"label":"minimum_effective_span_305","median_ms":1844.5498046875,"n":8,"spread_pct":15.45316147285685},{"features":{"max_depth":2,"packed_tokens":28799,"segment_count":6,"segments_below":[0,0,0,1,3,3,3,4]},"label":"minimum_effective_span_7015","median_ms":1883.3740234375,"n":8,"spread_pct":8.155788600391482},{"features":{"max_depth":2,"packed_tokens":37178,"segment_count":6,"segments_below":[0,0,0,1,2,2,2,3]},"label":"minimum_effective_span_8380","median_ms":2311.7154541015625,"n":8,"spread_pct":8.086506121983001},{"features":{"max_depth":1,"packed_tokens":48469,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":2842.17724609375,"n":8,"spread_pct":24.98953318791179},{"features":{"max_depth":3,"packed_tokens":22846,"segment_count":8,"segments_below":[0,0,0,2,4,6,6,7]},"label":"uniform_depth_2","median_ms":1864.807861328125,"n":8,"spread_pct":11.300718894752503},{"features":{"max_depth":4,"packed_tokens":21481,"segment_count":8,"segments_below":[0,0,0,2,5,6,6,7]},"label":"uniform_depth_3","median_ms":1893.6967163085938,"n":8,"spread_pct":41.534790760118014}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"62112dd684879b09adbef34aa98e5ea3966b7de9b9e8067de5920b1d73f07197","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":3,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":30477,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,2]},"label":"depth_one","median_ms":2030.5247802734375,"n":8,"spread_pct":1.630365745366179},{"features":{"max_depth":2,"packed_tokens":19821,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_495","median_ms":1621.454345703125,"n":8,"spread_pct":3.5242461880090827},{"features":{"max_depth":1,"packed_tokens":31465,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":2069.605224609375,"n":8,"spread_pct":0.7130030152989766},{"features":{"max_depth":3,"packed_tokens":19327,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,4]},"label":"uniform_depth_2","median_ms":1672.822998046875,"n":8,"spread_pct":1.5336368119537358}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"28b4492a5c1ce50958108e27b72b7afdf09941e4670076dfa879b517d9cc7575","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":4,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":13557,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1608.79248046875,"n":8,"spread_pct":46.49758696032878},{"features":{"max_depth":2,"packed_tokens":25902,"segment_count":6,"segments_below":[0,1,2,4,4,4,4,4]},"label":"minimum_effective_span_12406","median_ms":1793.8519897460938,"n":8,"spread_pct":1.9786004790464229},{"features":{"max_depth":2,"packed_tokens":37995,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12426","median_ms":2606.7310791015625,"n":8,"spread_pct":1.4516954655730185},{"features":{"max_depth":2,"packed_tokens":50582,"segment_count":6,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12592","median_ms":2840.476318359375,"n":8,"spread_pct":0.9774473378565693},{"features":{"max_depth":3,"packed_tokens":13367,"segment_count":7,"segments_below":[1,3,5,6,6,6,6,6]},"label":"minimum_effective_span_187","median_ms":1757.5579833984375,"n":8,"spread_pct":44.40238048356307},{"features":{"max_depth":3,"packed_tokens":13185,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_21","median_ms":1748.536376953125,"n":8,"spread_pct":3.5038093383068007},{"features":{"max_depth":1,"packed_tokens":63177,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":3684.4415283203125,"n":8,"spread_pct":18.770567285597878},{"features":{"max_depth":3,"packed_tokens":13497,"segment_count":7,"segments_below":[1,2,4,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1766.811767578125,"n":8,"spread_pct":1.0442303645185391},{"features":{"max_depth":4,"packed_tokens":13165,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1873.501953125,"n":8,"spread_pct":37.89214273278755},{"features":{"max_depth":5,"packed_tokens":13161,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1989.9535522460938,"n":8,"spread_pct":1.8393971745093205}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"7241823ac1b67d127c5921d771dda170dcc9d4886feffaff3c2520113cd01a08","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":5,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":9433,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1369.7730102539062,"n":8,"spread_pct":3.224928072407355},{"features":{"max_depth":6,"packed_tokens":6738,"segment_count":17,"segments_below":[6,11,15,16,16,16,16,17]},"label":"minimum_effective_span_105","median_ms":1880.9508056640625,"n":8,"spread_pct":39.280398428191674},{"features":{"max_depth":6,"packed_tokens":6778,"segment_count":17,"segments_below":[5,10,15,16,16,16,16,17]},"label":"minimum_effective_span_113","median_ms":1861.9218139648438,"n":8,"spread_pct":5.164151636182937},{"features":{"max_depth":6,"packed_tokens":6877,"segment_count":17,"segments_below":[5,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":1868.6859130859375,"n":8,"spread_pct":4.220638837491488},{"features":{"max_depth":5,"packed_tokens":7090,"segment_count":16,"segments_below":[3,8,14,15,15,15,15,16]},"label":"minimum_effective_span_164","median_ms":1753.825927734375,"n":8,"spread_pct":3.3668788562282432},{"features":{"max_depth":5,"packed_tokens":7134,"segment_count":16,"segments_below":[4,7,14,15,15,15,15,16]},"label":"minimum_effective_span_195","median_ms":1757.3363037109375,"n":8,"spread_pct":3.2098834552332547},{"features":{"max_depth":4,"packed_tokens":7415,"segment_count":16,"segments_below":[3,5,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":1658.9900512695312,"n":8,"spread_pct":47.72789473841054},{"features":{"max_depth":4,"packed_tokens":7626,"segment_count":15,"segments_below":[3,6,10,14,14,14,14,15]},"label":"minimum_effective_span_245","median_ms":1653.5840454101562,"n":8,"spread_pct":49.15142560137243},{"features":{"max_depth":4,"packed_tokens":7702,"segment_count":15,"segments_below":[3,5,9,14,14,14,14,15]},"label":"minimum_effective_span_281","median_ms":1647.720947265625,"n":8,"spread_pct":44.54525923220476},{"features":{"max_depth":12,"packed_tokens":6323,"segment_count":23,"segments_below":[16,20,22,22,22,22,22,23]},"label":"minimum_effective_span_32","median_ms":2570.962646484375,"n":8,"spread_pct":4.0321582791131005},{"features":{"max_depth":3,"packed_tokens":8041,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_340","median_ms":1538.6173095703125,"n":8,"spread_pct":4.196439119088461},{"features":{"max_depth":3,"packed_tokens":8113,"segment_count":15,"segments_below":[1,4,8,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":1532.11181640625,"n":8,"spread_pct":2.4508208868414303},{"features":{"max_depth":11,"packed_tokens":6359,"segment_count":22,"segments_below":[15,19,21,21,21,21,21,22]},"label":"minimum_effective_span_37","median_ms":2443.9541015625,"n":8,"spread_pct":2.6258596528662688},{"features":{"max_depth":9,"packed_tokens":6437,"segment_count":20,"segments_below":[9,17,19,19,19,19,19,20]},"label":"minimum_effective_span_40","median_ms":2234.2603759765625,"n":8,"spread_pct":30.546846781530153},{"features":{"max_depth":3,"packed_tokens":8391,"segment_count":15,"segments_below":[1,3,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":1529.71826171875,"n":8,"spread_pct":3.3524127930266983},{"features":{"max_depth":2,"packed_tokens":13958,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4955","median_ms":1501.651123046875,"n":8,"spread_pct":4.339087687169868},{"features":{"max_depth":2,"packed_tokens":18411,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4994","median_ms":1748.3311157226562,"n":8,"spread_pct":1.676508118075912},{"features":{"max_depth":2,"packed_tokens":23107,"segment_count":14,"segments_below":[2,3,5,7,10,10,10,14]},"label":"minimum_effective_span_5048","median_ms":1766.2793579101562,"n":8,"spread_pct":2.974996718065045},{"features":{"max_depth":2,"packed_tokens":27945,"segment_count":14,"segments_below":[2,3,5,6,9,9,9,14]},"label":"minimum_effective_span_5087","median_ms":2031.640380859375,"n":8,"spread_pct":35.766972188878604},{"features":{"max_depth":2,"packed_tokens":32845,"segment_count":14,"segments_below":[1,3,4,6,8,8,8,14]},"label":"minimum_effective_span_5118","median_ms":2080.94287109375,"n":8,"spread_pct":1.8345894230356348},{"features":{"max_depth":2,"packed_tokens":37693,"segment_count":14,"segments_below":[2,2,3,6,7,7,7,14]},"label":"minimum_effective_span_5149","median_ms":2346.6510009765625,"n":8,"spread_pct":1.7543909842842962},{"features":{"max_depth":2,"packed_tokens":42711,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_5199","median_ms":2402.557373046875,"n":8,"spread_pct":1.8789284489438185},{"features":{"max_depth":3,"packed_tokens":8746,"segment_count":14,"segments_below":[2,4,7,10,13,13,13,14]},"label":"minimum_effective_span_522","median_ms":1443.29541015625,"n":8,"spread_pct":54.610200925683785},{"features":{"max_depth":2,"packed_tokens":47673,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5235","median_ms":2710.97802734375,"n":8,"spread_pct":1.0555137594304698},{"features":{"max_depth":2,"packed_tokens":52699,"segment_count":14,"segments_below":[0,1,2,4,4,4,4,14]},"label":"minimum_effective_span_5303","median_ms":2803.577880859375,"n":8,"spread_pct":1.4752877074016486},{"features":{"max_depth":2,"packed_tokens":57931,"segment_count":14,"segments_below":[0,0,3,3,3,3,3,14]},"label":"minimum_effective_span_5395","median_ms":3132.416748046875,"n":8,"spread_pct":0.6291006680317082},{"features":{"max_depth":2,"packed_tokens":63240,"segment_count":13,"segments_below":[1,1,1,1,1,1,1,13]},"label":"minimum_effective_span_5476","median_ms":3149.5064697265625,"n":8,"spread_pct":0.3891203657319047},{"features":{"max_depth":8,"packed_tokens":6558,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_71","median_ms":2116.6025390625,"n":8,"spread_pct":4.077937123771743},{"features":{"max_depth":7,"packed_tokens":6639,"segment_count":18,"segments_below":[6,14,16,17,17,17,17,18]},"label":"minimum_effective_span_82","median_ms":1994.8646850585938,"n":8,"spread_pct":34.867004552062895},{"features":{"max_depth":7,"packed_tokens":6657,"segment_count":18,"segments_below":[6,14,16,17,17,17,17,18]},"label":"minimum_effective_span_87","median_ms":1981.1920166015625,"n":8,"spread_pct":5.0228389841604315},{"features":{"max_depth":7,"packed_tokens":6668,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_93","median_ms":2004.7822265625,"n":8,"spread_pct":40.13685294488715},{"features":{"max_depth":7,"packed_tokens":6645,"segment_count":18,"segments_below":[7,12,17,17,17,17,17,18]},"label":"minimum_effective_span_94","median_ms":1995.693115234375,"n":8,"spread_pct":3.65744885929282},{"features":{"max_depth":1,"packed_tokens":68881,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":3528.3037109375,"n":8,"spread_pct":19.552513809226337},{"features":{"max_depth":11,"packed_tokens":6620,"segment_count":23,"segments_below":[16,19,20,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":2454.2960205078125,"n":8,"spread_pct":28.494051381623013},{"features":{"max_depth":12,"packed_tokens":6458,"segment_count":24,"segments_below":[16,20,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":2583.0914306640625,"n":8,"spread_pct":31.588107851896098},{"features":{"max_depth":13,"packed_tokens":6292,"segment_count":24,"segments_below":[17,21,23,23,23,23,23,24]},"label":"uniform_depth_12","median_ms":2680.6624755859375,"n":8,"spread_pct":2.5485443138012784},{"features":{"max_depth":3,"packed_tokens":9004,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1534.291259765625,"n":8,"spread_pct":1.4418906836342424},{"features":{"max_depth":4,"packed_tokens":8464,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1641.3304443359375,"n":8,"spread_pct":2.3616854032778614},{"features":{"max_depth":5,"packed_tokens":8113,"segment_count":17,"segments_below":[8,9,11,13,16,16,16,17]},"label":"uniform_depth_4","median_ms":1772.9521484375,"n":8,"spread_pct":2.1031155002222524},{"features":{"max_depth":6,"packed_tokens":7865,"segment_count":18,"segments_below":[10,11,13,14,17,17,17,18]},"label":"uniform_depth_5","median_ms":1883.2142944335938,"n":8,"spread_pct":3.5604759267043082},{"features":{"max_depth":7,"packed_tokens":7648,"segment_count":19,"segments_below":[11,13,14,16,18,18,18,19]},"label":"uniform_depth_6","median_ms":2020.8404541015625,"n":8,"spread_pct":42.532853913308735},{"features":{"max_depth":8,"packed_tokens":7348,"segment_count":20,"segments_below":[14,14,15,18,19,19,19,20]},"label":"uniform_depth_7","median_ms":2130.344970703125,"n":8,"spread_pct":3.3263529589464365},{"features":{"max_depth":9,"packed_tokens":7168,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":2232.7825927734375,"n":8,"spread_pct":1.612906488480235},{"features":{"max_depth":10,"packed_tokens":6896,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":2337.8096923828125,"n":8,"spread_pct":32.95051410770956}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"22fe33a263cacb6bc86ab088fe07af0a0fb8e68bfdc85c518a57ec37e403d138","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":6,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":12659,"segment_count":5,"segments_below":[1,3,4,4,4,4,4,4]},"label":"depth_one","median_ms":1197.895751953125,"n":8,"spread_pct":5.607275208541345},{"features":{"max_depth":2,"packed_tokens":24836,"segment_count":5,"segments_below":[2,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12268","median_ms":1304.2882690429688,"n":8,"spread_pct":2.3060012890359087},{"features":{"max_depth":2,"packed_tokens":37143,"segment_count":5,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12313","median_ms":1469.8665161132812,"n":8,"spread_pct":51.90507398942451},{"features":{"max_depth":3,"packed_tokens":12609,"segment_count":6,"segments_below":[3,5,5,5,5,5,5,5]},"label":"minimum_effective_span_46","median_ms":1448.642822265625,"n":8,"spread_pct":5.366772273550085},{"features":{"max_depth":1,"packed_tokens":49460,"segment_count":4,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1624.3839111328125,"n":8,"spread_pct":50.187154038944904},{"features":{"max_depth":3,"packed_tokens":12569,"segment_count":6,"segments_below":[4,4,5,5,5,5,5,5]},"label":"uniform_depth_2","median_ms":1428.5100708007812,"n":8,"spread_pct":5.761960125203864},{"features":{"max_depth":4,"packed_tokens":12564,"segment_count":7,"segments_below":[5,6,6,6,6,6,6,6]},"label":"uniform_depth_3","median_ms":1583.042724609375,"n":8,"spread_pct":154.834732978591}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g0","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"f8b37d8f36ed38a4d2ea74f20bfe0d15bedbdb66ffaae41eb728a1756d67d5ff","shape":[1,4,4,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":0,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":12382,"segment_count":3,"segments_below":[1,2,2,2,2,2,2,2]},"label":"depth_one","median_ms":1176.4012451171875,"n":8,"spread_pct":61.49455659749184},{"features":{"max_depth":1,"packed_tokens":24645,"segment_count":2,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1289.9717407226562,"n":8,"spread_pct":56.25502693483673}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"515df421b93fe5f98e250ae3d849975dccadfd7dc00046242b5875ddf392c74f","shape":[1,4,4,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":1,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":1,"packed_tokens":12546,"segment_count":1,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":1183.51708984375,"n":8,"spread_pct":3.602605449280337}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"8aa85999c5d3c576d127bd2e3a44436cbd0e9062c7c92b16bcbaab0814703094","shape":[1,4,4,1],"source":"c8d2d7dc0aae7b0e64d6d11bc347c85ef0cca2c7b2391094c70bbececc4da349","workload":{"group":2,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":47253,"segment_count":6,"segments_below":[0,0,0,1,1,1,1,3]},"label":"depth_one","median_ms":1745.0889892578125,"n":8,"spread_pct":0.6376222223512426},{"features":{"max_depth":2,"packed_tokens":23150,"segment_count":7,"segments_below":[0,0,0,1,3,5,5,6]},"label":"minimum_effective_span_1366","median_ms":1254.4878540039062,"n":8,"spread_pct":1.3984946332087302},{"features":{"max_depth":3,"packed_tokens":21785,"segment_count":7,"segments_below":[0,0,0,1,4,5,5,6]},"label":"minimum_effective_span_305","median_ms":1257.8349609375,"n":8,"spread_pct":2.0543693707332866},{"features":{"max_depth":2,"packed_tokens":28799,"segment_count":6,"segments_below":[0,0,0,1,3,3,3,4]},"label":"minimum_effective_span_7015","median_ms":1292.5856323242188,"n":8,"spread_pct":1.811054002358988},{"features":{"max_depth":2,"packed_tokens":37178,"segment_count":6,"segments_below":[0,0,0,1,2,2,2,3]},"label":"minimum_effective_span_8380","median_ms":1368.750732421875,"n":8,"spread_pct":1.6963817977409046},{"features":{"max_depth":1,"packed_tokens":48469,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":1738.1285400390625,"n":8,"spread_pct":0.5707242960740442},{"features":{"max_depth":3,"packed_tokens":22846,"segment_count":8,"segments_below":[0,0,0,2,4,6,6,7]},"label":"uniform_depth_2","median_ms":1278.6812744140625,"n":8,"spread_pct":55.757364922834526},{"features":{"max_depth":4,"packed_tokens":21481,"segment_count":8,"segments_below":[0,0,0,2,5,6,6,7]},"label":"uniform_depth_3","median_ms":1280.7189331054688,"n":8,"spread_pct":111.1358990785958}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g3","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"62112dd684879b09adbef34aa98e5ea3966b7de9b9e8067de5920b1d73f07197","shape":[1,4,4,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":3,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":30477,"segment_count":4,"segments_below":[0,0,0,1,1,1,1,2]},"label":"depth_one","median_ms":1339.6716918945312,"n":8,"spread_pct":122.15118185958738},{"features":{"max_depth":2,"packed_tokens":19821,"segment_count":4,"segments_below":[0,0,0,2,2,2,2,3]},"label":"minimum_effective_span_495","median_ms":1218.4442138671875,"n":8,"spread_pct":3.104965739100205},{"features":{"max_depth":1,"packed_tokens":31465,"segment_count":3,"segments_below":[0,0,0,0,0,0,0,1]},"label":"no_sharing","median_ms":1360.9903564453125,"n":8,"spread_pct":0.8207739492035012},{"features":{"max_depth":3,"packed_tokens":19327,"segment_count":5,"segments_below":[0,0,0,3,3,3,3,4]},"label":"uniform_depth_2","median_ms":1224.7328491210938,"n":8,"spread_pct":59.87928948352147}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"28b4492a5c1ce50958108e27b72b7afdf09941e4670076dfa879b517d9cc7575","shape":[1,4,4,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":4,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":13557,"segment_count":6,"segments_below":[0,1,3,5,5,5,5,5]},"label":"depth_one","median_ms":1355.66796875,"n":8,"spread_pct":131.35580217316763},{"features":{"max_depth":2,"packed_tokens":25902,"segment_count":6,"segments_below":[0,1,2,4,4,4,4,4]},"label":"minimum_effective_span_12406","median_ms":1443.4220581054688,"n":8,"spread_pct":4.755823569400974},{"features":{"max_depth":2,"packed_tokens":37995,"segment_count":6,"segments_below":[1,2,3,3,3,3,3,3]},"label":"minimum_effective_span_12426","median_ms":1580.6881713867188,"n":8,"spread_pct":4.045494641573604},{"features":{"max_depth":2,"packed_tokens":50582,"segment_count":6,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_12592","median_ms":1787.9703369140625,"n":8,"spread_pct":3.6790745551340596},{"features":{"max_depth":3,"packed_tokens":13367,"segment_count":7,"segments_below":[1,3,5,6,6,6,6,6]},"label":"minimum_effective_span_187","median_ms":1611.227783203125,"n":8,"spread_pct":2.98637536367795},{"features":{"max_depth":3,"packed_tokens":13185,"segment_count":7,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_21","median_ms":1608.3685302734375,"n":8,"spread_pct":7.72590261430939},{"features":{"max_depth":1,"packed_tokens":63177,"segment_count":5,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2140.9090576171875,"n":8,"spread_pct":83.56914544927803},{"features":{"max_depth":3,"packed_tokens":13497,"segment_count":7,"segments_below":[1,2,4,6,6,6,6,6]},"label":"uniform_depth_2","median_ms":1597.83544921875,"n":8,"spread_pct":36.9058773053267},{"features":{"max_depth":4,"packed_tokens":13165,"segment_count":8,"segments_below":[2,4,7,7,7,7,7,7]},"label":"uniform_depth_3","median_ms":1738.8045654296875,"n":8,"spread_pct":42.67878764307403},{"features":{"max_depth":5,"packed_tokens":13161,"segment_count":9,"segments_below":[3,5,8,8,8,8,8,8]},"label":"uniform_depth_4","median_ms":1859.856689453125,"n":8,"spread_pct":38.37917251741375}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g5","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"7241823ac1b67d127c5921d771dda170dcc9d4886feffaff3c2520113cd01a08","shape":[1,4,4,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":5,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":9433,"segment_count":14,"segments_below":[1,3,6,9,13,13,13,14]},"label":"depth_one","median_ms":1001.00390625,"n":8,"spread_pct":2.4544878286251692},{"features":{"max_depth":6,"packed_tokens":6738,"segment_count":17,"segments_below":[6,11,15,16,16,16,16,17]},"label":"minimum_effective_span_105","median_ms":1489.1924438476562,"n":8,"spread_pct":3.7085808151661084},{"features":{"max_depth":6,"packed_tokens":6778,"segment_count":17,"segments_below":[5,10,15,16,16,16,16,17]},"label":"minimum_effective_span_113","median_ms":1510.5835571289062,"n":8,"spread_pct":50.10980265744613},{"features":{"max_depth":6,"packed_tokens":6877,"segment_count":17,"segments_below":[5,8,16,16,16,16,16,17]},"label":"minimum_effective_span_133","median_ms":1510.6831665039062,"n":8,"spread_pct":2.8494164263770325},{"features":{"max_depth":5,"packed_tokens":7090,"segment_count":16,"segments_below":[3,8,14,15,15,15,15,16]},"label":"minimum_effective_span_164","median_ms":1389.8375244140625,"n":8,"spread_pct":3.9057053814444957},{"features":{"max_depth":5,"packed_tokens":7134,"segment_count":16,"segments_below":[4,7,14,15,15,15,15,16]},"label":"minimum_effective_span_195","median_ms":1390.1932983398438,"n":8,"spread_pct":4.610780023134815},{"features":{"max_depth":4,"packed_tokens":7415,"segment_count":16,"segments_below":[3,5,13,15,15,15,15,16]},"label":"minimum_effective_span_197","median_ms":1272.5956420898438,"n":8,"spread_pct":3.9109347931174017},{"features":{"max_depth":4,"packed_tokens":7626,"segment_count":15,"segments_below":[3,6,10,14,14,14,14,15]},"label":"minimum_effective_span_245","median_ms":1179.1629638671875,"n":8,"spread_pct":62.911820373773395},{"features":{"max_depth":4,"packed_tokens":7702,"segment_count":15,"segments_below":[3,5,9,14,14,14,14,15]},"label":"minimum_effective_span_281","median_ms":1185.3734741210938,"n":8,"spread_pct":12.238764947035344},{"features":{"max_depth":12,"packed_tokens":6323,"segment_count":23,"segments_below":[16,20,22,22,22,22,22,23]},"label":"minimum_effective_span_32","median_ms":2235.8277587890625,"n":8,"spread_pct":3.897979138454614},{"features":{"max_depth":3,"packed_tokens":8041,"segment_count":15,"segments_below":[2,4,8,14,14,14,14,15]},"label":"minimum_effective_span_340","median_ms":1116.1997680664062,"n":8,"spread_pct":187.57786126140698},{"features":{"max_depth":3,"packed_tokens":8113,"segment_count":15,"segments_below":[1,4,8,14,14,14,14,15]},"label":"minimum_effective_span_349","median_ms":1112.76318359375,"n":8,"spread_pct":5.1938314276950805},{"features":{"max_depth":11,"packed_tokens":6359,"segment_count":22,"segments_below":[15,19,21,21,21,21,21,22]},"label":"minimum_effective_span_37","median_ms":2130.4166259765625,"n":8,"spread_pct":2.966163970939654},{"features":{"max_depth":9,"packed_tokens":6437,"segment_count":20,"segments_below":[9,17,19,19,19,19,19,20]},"label":"minimum_effective_span_40","median_ms":1872.2396850585938,"n":8,"spread_pct":4.039494992220806},{"features":{"max_depth":3,"packed_tokens":8391,"segment_count":15,"segments_below":[1,3,9,12,14,14,14,15]},"label":"minimum_effective_span_441","median_ms":1087.1493530273438,"n":8,"spread_pct":4.248395240136415},{"features":{"max_depth":2,"packed_tokens":13958,"segment_count":14,"segments_below":[1,3,6,9,12,12,12,14]},"label":"minimum_effective_span_4955","median_ms":1059.6612548828125,"n":8,"spread_pct":8.665961274061155},{"features":{"max_depth":2,"packed_tokens":18411,"segment_count":14,"segments_below":[2,4,6,8,11,11,11,14]},"label":"minimum_effective_span_4994","median_ms":1095.3551635742188,"n":8,"spread_pct":8.069652689950168},{"features":{"max_depth":2,"packed_tokens":23107,"segment_count":14,"segments_below":[2,3,5,7,10,10,10,14]},"label":"minimum_effective_span_5048","median_ms":1191.3729858398438,"n":8,"spread_pct":11.415513394200811},{"features":{"max_depth":2,"packed_tokens":27945,"segment_count":14,"segments_below":[2,3,5,6,9,9,9,14]},"label":"minimum_effective_span_5087","median_ms":1260.2797241210938,"n":8,"spread_pct":58.44427925561192},{"features":{"max_depth":2,"packed_tokens":32845,"segment_count":14,"segments_below":[1,3,4,6,8,8,8,14]},"label":"minimum_effective_span_5118","median_ms":1461.8350219726562,"n":8,"spread_pct":2.840868358572326},{"features":{"max_depth":2,"packed_tokens":37693,"segment_count":14,"segments_below":[2,2,3,6,7,7,7,14]},"label":"minimum_effective_span_5149","median_ms":1507.6705322265625,"n":8,"spread_pct":1.593244247333197},{"features":{"max_depth":2,"packed_tokens":42711,"segment_count":14,"segments_below":[1,2,3,6,6,6,6,14]},"label":"minimum_effective_span_5199","median_ms":1555.9444580078125,"n":8,"spread_pct":2.0906150233279517},{"features":{"max_depth":3,"packed_tokens":8746,"segment_count":14,"segments_below":[2,4,7,10,13,13,13,14]},"label":"minimum_effective_span_522","median_ms":1084.8883666992188,"n":8,"spread_pct":5.286289682630099},{"features":{"max_depth":2,"packed_tokens":47673,"segment_count":14,"segments_below":[1,1,2,5,5,5,5,14]},"label":"minimum_effective_span_5235","median_ms":1823.4406127929688,"n":8,"spread_pct":40.94646100211845},{"features":{"max_depth":2,"packed_tokens":52699,"segment_count":14,"segments_below":[0,1,2,4,4,4,4,14]},"label":"minimum_effective_span_5303","median_ms":1879.6024169921875,"n":8,"spread_pct":0.9996231256051633},{"features":{"max_depth":2,"packed_tokens":57931,"segment_count":14,"segments_below":[0,0,3,3,3,3,3,14]},"label":"minimum_effective_span_5395","median_ms":1925.2188720703125,"n":8,"spread_pct":36.07172175596664},{"features":{"max_depth":2,"packed_tokens":63240,"segment_count":13,"segments_below":[1,1,1,1,1,1,1,13]},"label":"minimum_effective_span_5476","median_ms":1903.5664672851562,"n":8,"spread_pct":38.89119265713231},{"features":{"max_depth":8,"packed_tokens":6558,"segment_count":19,"segments_below":[6,15,18,18,18,18,18,19]},"label":"minimum_effective_span_71","median_ms":1753.1758422851562,"n":8,"spread_pct":4.454555528466354},{"features":{"max_depth":7,"packed_tokens":6639,"segment_count":18,"segments_below":[6,14,16,17,17,17,17,18]},"label":"minimum_effective_span_82","median_ms":1616.23388671875,"n":8,"spread_pct":5.702698481320658},{"features":{"max_depth":7,"packed_tokens":6657,"segment_count":18,"segments_below":[6,14,16,17,17,17,17,18]},"label":"minimum_effective_span_87","median_ms":1630.106201171875,"n":8,"spread_pct":3.556095891590652},{"features":{"max_depth":7,"packed_tokens":6668,"segment_count":18,"segments_below":[6,12,17,17,17,17,17,18]},"label":"minimum_effective_span_93","median_ms":1632.484619140625,"n":8,"spread_pct":2.414689433000676},{"features":{"max_depth":7,"packed_tokens":6645,"segment_count":18,"segments_below":[7,12,17,17,17,17,17,18]},"label":"minimum_effective_span_94","median_ms":1627.2741088867188,"n":8,"spread_pct":4.757479757426309},{"features":{"max_depth":1,"packed_tokens":68881,"segment_count":13,"segments_below":[0,0,0,0,0,0,0,13]},"label":"no_sharing","median_ms":2167.6666259765625,"n":8,"spread_pct":63.49471562056086},{"features":{"max_depth":11,"packed_tokens":6620,"segment_count":23,"segments_below":[16,19,20,22,22,22,22,23]},"label":"uniform_depth_10","median_ms":2122.15966796875,"n":8,"spread_pct":1.7924003660223236},{"features":{"max_depth":12,"packed_tokens":6458,"segment_count":24,"segments_below":[16,20,23,23,23,23,23,24]},"label":"uniform_depth_11","median_ms":2241.8223876953125,"n":8,"spread_pct":32.929578199498835},{"features":{"max_depth":13,"packed_tokens":6292,"segment_count":24,"segments_below":[17,21,23,23,23,23,23,24]},"label":"uniform_depth_12","median_ms":2365.6507568359375,"n":8,"spread_pct":30.3009105901103},{"features":{"max_depth":3,"packed_tokens":9004,"segment_count":15,"segments_below":[3,5,8,11,14,14,14,15]},"label":"uniform_depth_2","median_ms":1235.8958129882812,"n":8,"spread_pct":59.36409339818026},{"features":{"max_depth":4,"packed_tokens":8464,"segment_count":16,"segments_below":[6,8,10,12,15,15,15,16]},"label":"uniform_depth_3","median_ms":1374.74169921875,"n":8,"spread_pct":7.898323229044093},{"features":{"max_depth":5,"packed_tokens":8113,"segment_count":17,"segments_below":[8,9,11,13,16,16,16,17]},"label":"uniform_depth_4","median_ms":1511.1358032226562,"n":8,"spread_pct":49.363972678579344},{"features":{"max_depth":6,"packed_tokens":7865,"segment_count":18,"segments_below":[10,11,13,14,17,17,17,18]},"label":"uniform_depth_5","median_ms":1532.4934692382812,"n":8,"spread_pct":4.723364592361281},{"features":{"max_depth":7,"packed_tokens":7648,"segment_count":19,"segments_below":[11,13,14,16,18,18,18,19]},"label":"uniform_depth_6","median_ms":1669.6234130859375,"n":8,"spread_pct":45.47429712401083},{"features":{"max_depth":8,"packed_tokens":7348,"segment_count":20,"segments_below":[14,14,15,18,19,19,19,20]},"label":"uniform_depth_7","median_ms":1803.7139282226562,"n":8,"spread_pct":2.773838118636038},{"features":{"max_depth":9,"packed_tokens":7168,"segment_count":21,"segments_below":[15,16,17,20,20,20,20,21]},"label":"uniform_depth_8","median_ms":1774.0901489257812,"n":8,"spread_pct":41.54440272926817},{"features":{"max_depth":10,"packed_tokens":6896,"segment_count":22,"segments_below":[16,17,18,21,21,21,21,22]},"label":"uniform_depth_9","median_ms":1865.8738403320312,"n":8,"spread_pct":42.56909697549044}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g6","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"22fe33a263cacb6bc86ab088fe07af0a0fb8e68bfdc85c518a57ec37e403d138","shape":[1,4,4,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":6,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":12674,"segment_count":9,"segments_below":[1,4,8,8,8,8,8,8]},"label":"depth_one","median_ms":1202.2567138671875,"n":8,"spread_pct":1.1989171974457622},{"features":{"max_depth":2,"packed_tokens":24329,"segment_count":9,"segments_below":[2,4,7,7,7,7,7,7]},"label":"minimum_effective_span_11722","median_ms":1271.0648803710938,"n":8,"spread_pct":0.9009308802774648},{"features":{"max_depth":2,"packed_tokens":36051,"segment_count":9,"segments_below":[1,3,6,6,6,6,6,6]},"label":"minimum_effective_span_11733","median_ms":1428.5572509765625,"n":8,"spread_pct":2.019338888214308},{"features":{"max_depth":2,"packed_tokens":47693,"segment_count":9,"segments_below":[1,2,5,5,5,5,5,5]},"label":"minimum_effective_span_11735","median_ms":1630.55859375,"n":8,"spread_pct":1.3439454701825246},{"features":{"max_depth":2,"packed_tokens":59300,"segment_count":9,"segments_below":[1,4,4,4,4,4,4,4]},"label":"minimum_effective_span_11758","median_ms":2408.6456298828125,"n":8,"spread_pct":0.5677687468720899},{"features":{"max_depth":2,"packed_tokens":71013,"segment_count":9,"segments_below":[2,3,3,3,3,3,3,3]},"label":"minimum_effective_span_11808","median_ms":2580.2274169921875,"n":8,"spread_pct":0.5326905998957384},{"features":{"max_depth":2,"packed_tokens":82862,"segment_count":9,"segments_below":[1,2,2,2,2,2,2,2]},"label":"minimum_effective_span_11855","median_ms":2751.874267578125,"n":8,"spread_pct":0.38515120687010457},{"features":{"max_depth":6,"packed_tokens":12273,"segment_count":13,"segments_below":[9,12,12,12,12,12,12,12]},"label":"minimum_effective_span_12","median_ms":1802.799072265625,"n":8,"spread_pct":39.66622270462937},{"features":{"max_depth":3,"packed_tokens":12536,"segment_count":10,"segments_below":[2,6,9,9,9,9,9,9]},"label":"minimum_effective_span_134","median_ms":1426.9893188476562,"n":8,"spread_pct":1.233475420419676},{"features":{"max_depth":5,"packed_tokens":12286,"segment_count":12,"segments_below":[8,11,11,11,11,11,11,11]},"label":"minimum_effective_span_14","median_ms":1670.2174072265625,"n":8,"spread_pct":1.8331624205979336},{"features":{"max_depth":7,"packed_tokens":12259,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"minimum_effective_span_3","median_ms":1909.5582275390625,"n":8,"spread_pct":2.12617690060302},{"features":{"max_depth":4,"packed_tokens":12322,"segment_count":11,"segments_below":[5,10,10,10,10,10,10,10]},"label":"minimum_effective_span_37","median_ms":1551.019775390625,"n":8,"spread_pct":1.6591355517263797},{"features":{"max_depth":4,"packed_tokens":12364,"segment_count":11,"segments_below":[4,10,10,10,10,10,10,10]},"label":"minimum_effective_span_48","median_ms":1548.5294799804688,"n":8,"spread_pct":48.10628089661204},{"features":{"max_depth":3,"packed_tokens":12416,"segment_count":10,"segments_below":[2,9,9,9,9,9,9,9]},"label":"minimum_effective_span_53","median_ms":1438.8058471679688,"n":8,"spread_pct":49.074993228595204},{"features":{"max_depth":6,"packed_tokens":12264,"segment_count":13,"segments_below":[10,12,12,12,12,12,12,12]},"label":"minimum_effective_span_6","median_ms":1805.4767456054688,"n":8,"spread_pct":2.46959772834736},{"features":{"max_depth":3,"packed_tokens":12408,"segment_count":10,"segments_below":[3,7,9,9,9,9,9,9]},"label":"minimum_effective_span_87","median_ms":1424.78271484375,"n":8,"spread_pct":1.9703696595732285},{"features":{"max_depth":1,"packed_tokens":94721,"segment_count":8,"segments_below":[0,0,0,0,0,0,0,0]},"label":"no_sharing","median_ms":2849.688720703125,"n":8,"spread_pct":0.4099611688331159},{"features":{"max_depth":3,"packed_tokens":12608,"segment_count":10,"segments_below":[4,6,9,9,9,9,9,9]},"label":"uniform_depth_2","median_ms":1428.1500244140625,"n":8,"spread_pct":2.622706026503057},{"features":{"max_depth":4,"packed_tokens":12598,"segment_count":11,"segments_below":[5,7,10,10,10,10,10,10]},"label":"uniform_depth_3","median_ms":1549.73828125,"n":8,"spread_pct":1.162368645915515},{"features":{"max_depth":5,"packed_tokens":12506,"segment_count":12,"segments_below":[6,8,11,11,11,11,11,11]},"label":"uniform_depth_4","median_ms":1671.6279907226562,"n":8,"spread_pct":69.82295463994699},{"features":{"max_depth":6,"packed_tokens":12356,"segment_count":13,"segments_below":[8,12,12,12,12,12,12,12]},"label":"uniform_depth_5","median_ms":1793.5,"n":8,"spread_pct":2.3969613578547535},{"features":{"max_depth":7,"packed_tokens":12262,"segment_count":14,"segments_below":[11,13,13,13,13,13,13,13]},"label":"uniform_depth_6","median_ms":1923.1084594726562,"n":8,"spread_pct":2.6765469278180807},{"features":{"max_depth":8,"packed_tokens":12257,"segment_count":15,"segments_below":[12,14,14,14,14,14,14,14]},"label":"uniform_depth_7","median_ms":2032.1137084960938,"n":8,"spread_pct":2.5846702022826773}],"cell":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g7","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"d04b77db2ed7d695d20c6f183215ead28d0337ae88b9ddbc3674926310d364b9","shape":[1,4,4,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"group":7,"kind":"ellavox"}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":3719.9403076171875,"n":8,"spread_pct":0.523329077539552},{"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1398.0547485351562,"n":8,"spread_pct":8.486750192495512},{"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":4896.9345703125,"n":8,"spread_pct":0.24208982780003674},{"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1525.22705078125,"n":8,"spread_pct":52.81762769054941}],"cell":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":3807.1978759765625,"n":8,"spread_pct":47.099762364828266},{"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1206.5172729492188,"n":8,"spread_pct":5.093764896551065},{"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":5050.154052734375,"n":8,"spread_pct":0.2741059637577771},{"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1345.854736328125,"n":8,"spread_pct":49.18887041209617}],"cell":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":2197.1409912109375,"n":8,"spread_pct":48.59830104807612},{"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1708.4969482421875,"n":8,"spread_pct":16.558515582337165},{"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":2686.1566162109375,"n":8,"spread_pct":5.413170964137915},{"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1748.7982788085938,"n":8,"spread_pct":19.0076814397275}],"cell":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":2063.1634521484375,"n":8,"spread_pct":1.0111922449188486},{"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1279.2296752929688,"n":8,"spread_pct":8.448134608597522},{"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":2623.1375732421875,"n":8,"spread_pct":26.194189208740593},{"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1314.9177856445312,"n":8,"spread_pct":5.272829585269657}],"cell":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,4,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":2243.066650390625,"n":8,"spread_pct":0.6519220849925894},{"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1114.6611328125,"n":8,"spread_pct":61.24170217090571},{"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":2926.552490234375,"n":8,"spread_pct":0.10777366013081165},{"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1132.5115356445312,"n":8,"spread_pct":3.469208333313123}],"cell":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[1,4,4,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":4218.9716796875,"n":8,"spread_pct":16.400700703921345},{"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1584.8867797851562,"n":8,"spread_pct":1.0782389887696917},{"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":5527.9208984375,"n":8,"spread_pct":0.3169280380794161},{"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1686.0945434570312,"n":8,"spread_pct":3.2740882678871754}],"cell":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":74719,"segment_count":33,"segments_below":[0,0,0,0,1,1,33,33]},"label":"depth_one","median_ms":4426.760498046875,"n":8,"spread_pct":14.805031830170408},{"features":{"max_depth":2,"packed_tokens":14302,"segment_count":34,"segments_below":[0,0,32,32,32,32,34,34]},"label":"minimum_effective_span_1024","median_ms":1364.6596069335938,"n":8,"spread_pct":54.05370277713932},{"features":{"max_depth":1,"packed_tokens":106432,"segment_count":32,"segments_below":[0,0,0,0,0,0,32,32]},"label":"no_sharing","median_ms":5847.463623046875,"n":8,"spread_pct":0.47657715511754295},{"features":{"max_depth":3,"packed_tokens":13279,"segment_count":35,"segments_below":[0,0,32,32,33,33,35,35]},"label":"uniform_depth_2","median_ms":1442.5343627929688,"n":8,"spread_pct":3.577772359141514}],"cell":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"e93bd99b9b146f0288d949abc36fd4bcf4c2ba5286f8e35dcf9a1b97782a7986","shape":[2,1,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6401,"shape":[2,16,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3010.2109375,"n":8,"spread_pct":28.59792645864207},{"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1722.6849975585938,"n":8,"spread_pct":45.10870726328228},{"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":3469.7777099609375,"n":8,"spread_pct":0.3287103247387396},{"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1701.7667236328125,"n":8,"spread_pct":2.702872338052287}],"cell":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3052.8104248046875,"n":8,"spread_pct":21.128184466613774},{"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1578.3125610351562,"n":8,"spread_pct":208.82852219799216},{"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":3567.2705078125,"n":8,"spread_pct":0.25307394018700285},{"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1541.4124755859375,"n":8,"spread_pct":44.87584827449775}],"cell":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1881.3087768554688,"n":8,"spread_pct":43.57242264577035},{"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1590.004638671875,"n":8,"spread_pct":12.447864996463043},{"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":2051.952392578125,"n":8,"spread_pct":3.5385268674543835},{"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1665.4912719726562,"n":8,"spread_pct":104.41165740580207}],"cell":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1699.6204833984375,"n":8,"spread_pct":40.712005516511866},{"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1165.1912841796875,"n":8,"spread_pct":5.6205774455076964},{"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":1926.1522827148438,"n":8,"spread_pct":1.0375219606998258},{"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1277.6146240234375,"n":8,"spread_pct":3.596819443583208}],"cell":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,4,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1814.5687255859375,"n":8,"spread_pct":37.72531445958065},{"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1088.9392700195312,"n":8,"spread_pct":4.685014966217763},{"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":2092.193603515625,"n":8,"spread_pct":0.775098093557186},{"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1232.7645263671875,"n":8,"spread_pct":4.700565126157971}],"cell":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[1,4,4,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3387.98974609375,"n":8,"spread_pct":0.5100234153762653},{"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1938.6181640625,"n":8,"spread_pct":1.3999481146213502},{"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":3921.951171875,"n":8,"spread_pct":0.39798778808578916},{"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1884.7791748046875,"n":8,"spread_pct":45.67765089253568}],"cell":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":58351,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3523.57275390625,"n":8,"spread_pct":1.1094363478563807},{"features":{"max_depth":2,"packed_tokens":24556,"segment_count":20,"segments_below":[0,0,0,16,16,16,20,20]},"label":"minimum_effective_span_1024","median_ms":1824.6738891601562,"n":8,"spread_pct":41.822254130451995},{"features":{"max_depth":1,"packed_tokens":73696,"segment_count":16,"segments_below":[0,0,0,0,0,0,0,16]},"label":"no_sharing","median_ms":4144.834716796875,"n":8,"spread_pct":1.3866680237774769},{"features":{"max_depth":3,"packed_tokens":21487,"segment_count":21,"segments_below":[0,0,0,16,17,17,21,21]},"label":"uniform_depth_2","median_ms":1751.9749755859375,"n":8,"spread_pct":4.438998426650058}],"cell":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"b22cf02205e50cf3039bce79a80192ff93bea622b889e06740c68439c9ad1dc4","shape":[2,1,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6501,"shape":[4,4,1023,3072,511]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":3625.0157470703125,"n":8,"spread_pct":1.2222468457084086},{"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1668.4222412109375,"n":8,"spread_pct":3.263552100698617},{"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":4575.041015625,"n":8,"spread_pct":0.9661478444737304},{"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1655.1027221679688,"n":8,"spread_pct":2.193662819871209}],"cell":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":2621.5723876953125,"n":8,"spread_pct":47.16686190082372},{"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":2212.41796875,"n":8,"spread_pct":44.11396233793979},{"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":2868.099609375,"n":8,"spread_pct":14.101304285353363},{"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":2406.951904296875,"n":8,"spread_pct":46.42721405839916}],"cell":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":2169.6103515625,"n":8,"spread_pct":41.64271333283221},{"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1267.69921875,"n":8,"spread_pct":4.829835208346815},{"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":2660.2635498046875,"n":8,"spread_pct":30.386980995030534},{"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1404.980224609375,"n":8,"spread_pct":6.133072588634286}],"cell":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1474.8010864257812,"n":8,"spread_pct":55.736752105200225},{"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1335.1529541015625,"n":8,"spread_pct":69.25088408552793},{"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":1606.5645141601562,"n":8,"spread_pct":3.142441596183991},{"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1427.5499267578125,"n":8,"spread_pct":2.668863940955436}],"cell":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1282.3115234375,"n":8,"spread_pct":2.038783416483447},{"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1130.4710083007812,"n":8,"spread_pct":69.13474224767498},{"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":1464.0324096679688,"n":8,"spread_pct":47.799516623936746},{"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1228.3031616210938,"n":8,"spread_pct":66.2261890582329}],"cell":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,4,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":1272.5842895507812,"n":8,"spread_pct":1.6524683596241942},{"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1057.3104248046875,"n":8,"spread_pct":4.206390888770023},{"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":1533.196044921875,"n":8,"spread_pct":0.8682285316285697},{"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1113.0418090820312,"n":8,"spread_pct":4.4307446168383695}],"cell":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[1,4,4,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":2618.22412109375,"n":8,"spread_pct":3.7437317271535915},{"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1664.0608520507812,"n":8,"spread_pct":5.8105911416505585},{"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":3082.9234619140625,"n":8,"spread_pct":1.5349089571622347},{"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1864.1677856445312,"n":8,"spread_pct":5.625343680655086}],"cell":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":37871,"segment_count":17,"segments_below":[0,0,0,0,1,1,17,17]},"label":"depth_one","median_ms":2492.1180419921875,"n":8,"spread_pct":1.9070280710703584},{"features":{"max_depth":2,"packed_tokens":10222,"segment_count":18,"segments_below":[0,0,16,16,16,16,18,18]},"label":"minimum_effective_span_1024","median_ms":1404.07470703125,"n":8,"spread_pct":62.815561919566555},{"features":{"max_depth":1,"packed_tokens":53216,"segment_count":16,"segments_below":[0,0,0,0,0,0,16,16]},"label":"no_sharing","median_ms":3108.6719970703125,"n":8,"spread_pct":1.8831204809535886},{"features":{"max_depth":3,"packed_tokens":9199,"segment_count":19,"segments_below":[0,0,16,16,17,17,19,19]},"label":"uniform_depth_2","median_ms":1630.5601806640625,"n":8,"spread_pct":7.369989808759077}],"cell":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"89962c92661d053c96a70ffd9e4618ef80caa27f883ab722e8ac094e9530fc2b","shape":[2,1,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"grpo","seed":6301,"shape":[2,8,1023,2048,255]}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":2821.97705078125,"n":8,"spread_pct":0.452857722995552},{"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":3201.9176025390625,"n":8,"spread_pct":26.949312334157664},{"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":2860.3858642578125,"n":8,"spread_pct":0.6893908450483331},{"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":3451.397216796875,"n":8,"spread_pct":0.328897596740111},{"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":4196.81591796875,"n":8,"spread_pct":0.5011355916565746}],"cell":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":2150.674560546875,"n":8,"spread_pct":3.851137730111249},{"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":2455.2568359375,"n":8,"spread_pct":2.2046326778917846},{"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":2344.3441162109375,"n":8,"spread_pct":2.815203042020546},{"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":2410.4493408203125,"n":8,"spread_pct":3.144562900348271},{"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":2963.5975341796875,"n":8,"spread_pct":2.578084637107223}],"cell":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":1910.0067749023438,"n":8,"spread_pct":2.353328326337909},{"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":2189.392822265625,"n":8,"spread_pct":3.6107372716443016},{"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":2028.3070678710938,"n":8,"spread_pct":5.896072443660335},{"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":2171.1134033203125,"n":8,"spread_pct":36.37989962869736},{"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":2768.9002685546875,"n":8,"spread_pct":3.7394714971865723}],"cell":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":1581.0816650390625,"n":8,"spread_pct":4.4485266272093416},{"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":1614.5183715820312,"n":8,"spread_pct":4.876617578628132},{"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":1602.6521606445312,"n":8,"spread_pct":45.69353023645917},{"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":1650.6708374023438,"n":8,"spread_pct":2.707565442568534},{"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":1652.6241455078125,"n":8,"spread_pct":44.04501541072707}],"cell":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":1384.5230102539062,"n":8,"spread_pct":2.408708331083218},{"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":1415.970703125,"n":8,"spread_pct":52.91527351902199},{"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":1411.9022216796875,"n":8,"spread_pct":2.1414621422247024},{"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":1478.861572265625,"n":8,"spread_pct":45.120081711396125},{"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":1545.1627807617188,"n":8,"spread_pct":0.5335925217413356}],"cell":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,4,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":1377.5664672851562,"n":8,"spread_pct":50.39769297798157},{"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":1437.4282836914062,"n":8,"spread_pct":3.896551394696364},{"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":1392.7921142578125,"n":8,"spread_pct":11.902343547664879},{"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":1494.6245727539062,"n":8,"spread_pct":4.5897878910054},{"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":1607.1156005859375,"n":8,"spread_pct":0.7365239929145065}],"cell":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[1,4,4,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":2163.4283447265625,"n":8,"spread_pct":1.8083367728753998},{"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":2347.7586669921875,"n":8,"spread_pct":1.9976769940204875},{"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":2187.5693359375,"n":8,"spread_pct":33.4155145082455},{"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":2440.013671875,"n":8,"spread_pct":1.492111100171538},{"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":2717.34619140625,"n":8,"spread_pct":1.168922530940455}],"cell":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":28852,"segment_count":20,"segments_below":[0,0,1,4,9,14,19,20]},"label":"depth_one","median_ms":2107.087158203125,"n":8,"spread_pct":3.9686910421013346},{"features":{"max_depth":2,"packed_tokens":33460,"segment_count":18,"segments_below":[0,0,1,1,5,11,17,18]},"label":"minimum_effective_span_1025","median_ms":2334.6805419921875,"n":8,"spread_pct":1.8522508688961654},{"features":{"max_depth":2,"packed_tokens":29364,"segment_count":19,"segments_below":[0,0,1,3,8,13,18,19]},"label":"minimum_effective_span_257","median_ms":2137.0230712890625,"n":8,"spread_pct":34.027751874690146},{"features":{"max_depth":2,"packed_tokens":36532,"segment_count":17,"segments_below":[0,0,0,0,3,9,16,17]},"label":"minimum_effective_span_3073","median_ms":2462.849365234375,"n":8,"spread_pct":2.0196611144655376},{"features":{"max_depth":1,"packed_tokens":48820,"segment_count":16,"segments_below":[0,0,0,0,2,7,13,15]},"label":"no_sharing","median_ms":2830.9368896484375,"n":8,"spread_pct":0.6587200614375339}],"cell":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"1ce54a7d465ae286c4211f68951020c14d79bedee19dc7273c01c63a10524da0","shape":[2,1,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero2","seed":7778}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1937.2265625,"n":8,"spread_pct":2.1588291190288955},{"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1944.4967041015625,"n":8,"spread_pct":1.2543852319594115},{"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":2239.8162841796875,"n":8,"spread_pct":0.9798363088141723}],"cell":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1652.3369140625,"n":8,"spread_pct":57.64689186066373},{"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1591.710205078125,"n":8,"spread_pct":11.738445881639262},{"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1575.0780029296875,"n":8,"spread_pct":8.142244629346623}],"cell":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1270.4341430664062,"n":8,"spread_pct":1.6202226945038563},{"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1292.5760498046875,"n":8,"spread_pct":3.181073783455568},{"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1388.4468994140625,"n":8,"spread_pct":0.6256375747231201}],"cell":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1409.0783081054688,"n":8,"spread_pct":3.927179619081224},{"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1391.2659912109375,"n":8,"spread_pct":103.02097409278781},{"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1289.658203125,"n":8,"spread_pct":2.0837813586145795}],"cell":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1224.9118041992188,"n":8,"spread_pct":62.81958378002243},{"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1210.7443237304688,"n":8,"spread_pct":8.340584574101664},{"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1070.6791381835938,"n":8,"spread_pct":4.043665015069256}],"cell":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,4,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1104.695068359375,"n":8,"spread_pct":10.41944991978706},{"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1094.2777709960938,"n":8,"spread_pct":207.80853122937833},{"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1010.1430053710938,"n":8,"spread_pct":98.41741326969013}],"cell":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[1,4,4,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1701.4793701171875,"n":8,"spread_pct":2.6966853508611637},{"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1683.42041015625,"n":8,"spread_pct":2.203047297503078},{"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1748.7437744140625,"n":8,"spread_pct":1.556837431222427}],"cell":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":16756,"segment_count":13,"segments_below":[0,0,0,3,6,10,13,13]},"label":"depth_one","median_ms":1550.3907470703125,"n":8,"spread_pct":5.659217871376774},{"features":{"max_depth":2,"packed_tokens":17524,"segment_count":12,"segments_below":[0,0,0,2,4,8,12,12]},"label":"minimum_effective_span_769","median_ms":1516.9785766601562,"n":8,"spread_pct":7.694931922308271},{"features":{"max_depth":1,"packed_tokens":25204,"segment_count":11,"segments_below":[0,0,0,0,0,4,11,11]},"label":"no_sharing","median_ms":1643.0146484375,"n":8,"spread_pct":2.4453370596518047}],"cell":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"857a5db81b99dffffc7ddb02d60464dae255275fc94e22e4f96eb7164db117a4","shape":[2,1,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero3","seed":7779}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":2292.758056640625,"n":8,"spread_pct":1.8971398244962245},{"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":2625.8011474609375,"n":8,"spread_pct":1.4848053953771456},{"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":2382.771484375,"n":8,"spread_pct":29.590140486024758},{"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":2661.2467041015625,"n":8,"spread_pct":1.6809005376054318}],"cell":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1599.5739135742188,"n":8,"spread_pct":8.389929584443244},{"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1647.6054077148438,"n":8,"spread_pct":32.591147627206816},{"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1632.7681884765625,"n":8,"spread_pct":12.561584917413654},{"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1660.5042114257812,"n":8,"spread_pct":2.1850126192300654}],"cell":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,2,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1594.9544677734375,"n":8,"spread_pct":17.01804403795018},{"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1688.958984375,"n":8,"spread_pct":7.40561979399755},{"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1606.4295654296875,"n":8,"spread_pct":8.594871484682487},{"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1676.0845947265625,"n":8,"spread_pct":103.30375074150695}],"cell":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":2.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,2,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1480.7233276367188,"n":8,"spread_pct":49.98395930515082},{"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1441.2230224609375,"n":8,"spread_pct":51.131626742186654},{"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1429.6024780273438,"n":8,"spread_pct":51.20242166040697},{"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1338.3369750976562,"n":8,"spread_pct":4.566405377621964}],"cell":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,4,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1262.615478515625,"n":8,"spread_pct":57.44694221773109},{"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1192.7777709960938,"n":8,"spread_pct":2.7204236352710973},{"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1214.8496704101562,"n":8,"spread_pct":5.6557099776995665},{"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1107.8307495117188,"n":8,"spread_pct":69.63967613588876}],"cell":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,4,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1197.1098022460938,"n":8,"spread_pct":3.561719671203129},{"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":1170.8858032226562,"n":8,"spread_pct":3.781514948581858},{"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1132.265869140625,"n":8,"spread_pct":2.7403956348763985},{"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1108.3036499023438,"n":8,"spread_pct":5.924174022150683}],"cell":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":4.0,"ep":4.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":1.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[1,4,4,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":2155.3851318359375,"n":8,"spread_pct":37.42073856050165},{"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":2302.4488525390625,"n":8,"spread_pct":31.9920180942707},{"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":2191.556884765625,"n":8,"spread_pct":6.173795762924941},{"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":2237.4884033203125,"n":8,"spread_pct":32.372603624593246}],"cell":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":1.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[2,1,1,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}}, + {"candidates":[{"features":{"max_depth":2,"packed_tokens":19380,"segment_count":15,"segments_below":[0,0,1,4,8,10,14,15]},"label":"depth_one","median_ms":1802.9375610351562,"n":8,"spread_pct":5.20862203742001},{"features":{"max_depth":2,"packed_tokens":25012,"segment_count":13,"segments_below":[0,0,0,1,4,7,11,13]},"label":"minimum_effective_span_2049","median_ms":2013.4418334960938,"n":8,"spread_pct":4.034493392507164},{"features":{"max_depth":2,"packed_tokens":20916,"segment_count":14,"segments_below":[0,0,1,2,6,9,13,14]},"label":"minimum_effective_span_513","median_ms":1868.924072265625,"n":8,"spread_pct":5.267309574659446},{"features":{"max_depth":1,"packed_tokens":29108,"segment_count":12,"segments_below":[0,0,0,0,2,5,9,12]},"label":"no_sharing","median_ms":1934.1234130859375,"n":8,"spread_pct":1.5734326310488558}],"cell":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","device":"NVIDIA H200","device_capability":[9,0],"device_memory_class":"hbm-141g","facts":{"cp":1.0,"ep":2.0,"etp":1.0,"gdn_layers":30.0,"layers":40.0,"tp":2.0,"uses_gdn":1.0},"geometry":{"ffn_hidden_size":8192,"gdn_conv_kernel":4,"gdn_key_head_dim":128,"gdn_key_heads":16,"gdn_value_head_dim":128,"gdn_value_heads":32,"hidden_size":2048,"kv_channels":256,"moe_experts":256,"moe_ffn_hidden_size":512,"moe_shared_expert_ffn":512,"moe_topk":8,"num_attention_heads":16,"num_query_groups":2},"hidden_size":2048,"model":"Qwen/Qwen3.5-35B-A3B","param_dtype":"torch.bfloat16","requests_sha256":"ff3a1fcb7cc64401a9b7c23a99af2568e823fe915b381223e7fc8723b667a319","shape":[2,1,2,1],"source":"3ac077c7f400d7078e8c9a79828a74ba7e4d204b85b8dc9c112ca7a8e47480ee","workload":{"kind":"cal-hetero","seed":7777}} + ], + "fit_arguments": { + "exclude_cells": ["cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g5","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g6","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g7","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g2","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g7","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g0","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g1","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g2","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g3","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g4","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g5","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g6","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g7"], + "holdout": "", + "objective": "regret", + "require_complete": 8, + "terms": ["token_per_rank","token_cp_exchange","token_tp_collective","gdn_token_per_rank","attention_token_cp_exchange","tiny_segment_per_layer","level_cp_per_layer","level_tp_per_layer","gdn_level","gdn_level_tp"] + }, + "integer_table_milli_us": {"attention_token_cp_exchange":37,"gdn_level":8982505,"gdn_level_tp":0,"gdn_token_per_rank":368,"level_cp_per_layer":50356,"level_tp_per_layer":5701857,"tiny_segment_per_layer":442444,"token_cp_exchange":37,"token_per_rank":3145,"token_tp_collective":734}, + "integer_table_sha256": "eac01deb2f9cc160aaea78a31e9b870b32e70e68e2ab651d477431c98f635560", + "manifest": { + "excluded": ["cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g5","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g6","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g7","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g2","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g7","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g0","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g1","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g2","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g3","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g4","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g5","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g6","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g7","cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1"], + "expected": ["cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g0","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g1","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g2","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g3","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g4","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g5","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g6","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g7","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g0","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g1","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g2","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g3","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g4","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g5","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g6","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g7","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g0","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g1","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g2","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g3","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g4","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g5","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g6","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g7","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g0","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g1","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g2","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g3","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g4","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g5","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g6","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g7","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g0","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g1","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g2","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g3","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g4","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g5","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g6","cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g7","cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2"], + "path": "trainer_rank_cost_calibration_manifest_gdn-moe-h2048-h200-bf16.json" + }, + "measured_envelope": {"device_names":["NVIDIA H200"],"hidden_sizes":[2048],"param_dtypes":["torch.bfloat16"]}, + "metrics": { + "integer_all": {"cells":73,"clear_misses":[],"max_regret_pct":4.905084763364662,"median_regret_pct":0.0,"ordered_pairs":3658,"p95_regret_pct":2.3939137663455017,"pairwise_accuracy":0.972936030617824}, + "test": {"cells":13,"clear_misses":[],"max_regret_pct":0.8992007766222098,"median_regret_pct":0.0,"ordered_pairs":669,"p95_regret_pct":0.8992007766222098,"pairwise_accuracy":0.9596412556053812}, + "train": {"cells":60,"clear_misses":[],"max_regret_pct":4.905084763364662,"median_regret_pct":0.0,"ordered_pairs":2989,"p95_regret_pct":2.856514438270554,"pairwise_accuracy":0.9759116761458682} + }, + "schema": "art.dev.trainer_rank_cost_calibration_certificate.v1", + "table_id": "gdn-moe-h2048-h200-bf16" +} diff --git a/dev/trainer_rank_cost_calibration_lattice.sky.yaml b/dev/trainer_rank_cost_calibration_lattice.sky.yaml new file mode 100644 index 000000000..634601a6b --- /dev/null +++ b/dev/trainer_rank_cost_calibration_lattice.sky.yaml @@ -0,0 +1,105 @@ +# Cost-model calibration on a 4x H200 node over an explicit lattice of +# parallel shapes (TP, CP, EP[, ETP]) for one model: the shapes are run one +# after another with the matching torchrun world size (TP x CP; DP is 1, so +# EP must divide TP x CP). Used for the multi-class campaign: MoE models +# need EP isolated from CP (CP1/EP1, CP2/EP1, CP2/EP2, CP4/EP1, CP4/EP2, +# CP4/EP4) and dense controls reuse the same loop at EP1. +# +# The Ellavox corpora are git-ignored (customer-derived tokens); both are +# mounted so `cal-ellavox` (Qwen3.5 tokenizer) and `cal-ellavox-qwen3` +# (Qwen3 tokenizer) cells can run, and never committed. +# +# Usage: +# sky launch -c tr-cost-lattice dev/trainer_rank_cost_calibration_lattice.sky.yaml \ +# --env RUN_ID= --env MODEL=Qwen/Qwen3.5-35B-A3B \ +# --env SHAPES="1,1,1 1,2,1 1,2,2 1,4,1 1,4,2 1,4,4 2,1,1 2,1,2" \ +# --env CELLS="cal-grpo-g8:0 cal-ellavox:0:1" --env REPEAT=8 +# CELLS entries are cell:layers[:group]; layers 0 = full model. + +name: trainer-rank-cost-calibration-lattice + +workdir: . + +file_mounts: + /tmp/trainer_rank_ellavox_corpus.json: dev/_trainer_rank_ellavox_qwen35_4b_tokens.json + /tmp/trainer_rank_ellavox_qwen3_corpus.json: dev/_trainer_rank_ellavox_qwen3_06b_tokens.json + +envs: + RUN_ID: unset + MODEL: Qwen/Qwen3.5-35B-A3B + REPEAT: "8" + # tp,cp,ep[,etp] per shape; world size = tp x cp. + SHAPES: "1,1,1 1,2,1 1,2,2 1,4,1 1,4,2 1,4,4 2,1,1 2,1,2" + # cell:layers[:group] + CELLS: "cal-grpo-g8:0" + +resources: + cloud: kubernetes + accelerators: H200:4 + cpus: 32+ + memory: 256+ + image_id: docker:docker.io/bradhiltonnw/art-gpu@sha256:f93fda4032bf37f16d9c639c51daf73e1a884df97d14baabf349868831321bdc + +setup: | + set -euo pipefail + export CC=gcc CXX=g++ + uv sync --frozen --extra megatron --group dev + uv sync --project megatron_runtime --extra cuda12 --group test --frozen + uv pip install --python megatron_runtime/.venv/bin/python --no-deps --editable . + # Expert-parallel shapes need the HybridEP dispatcher, which ART builds from + # its vendored source for this node's GPU architecture (single node). + HYBRID_EP_MULTINODE=0 USE_NIXL=0 megatron_runtime/.venv/bin/python -m art.megatron.hybrid_ep_setup + +run: | + set -euo pipefail + : "${RUN_ID:?RUN_ID is required}" + root="scratch/trainer_rank_cost_calibration/${RUN_ID}" + mkdir -p "${root}" + cp /tmp/trainer_rank_ellavox_corpus.json dev/_trainer_rank_ellavox_qwen35_4b_tokens.json + cp /tmp/trainer_rank_ellavox_qwen3_corpus.json dev/_trainer_rank_ellavox_qwen3_06b_tokens.json + export ART_MEGATRON_DATA_PARALLEL_SIZE=1 + export ART_MEGATRON_PIPELINE_MODEL_PARALLEL_SIZE=1 + export TOKENIZERS_PARALLELISM=false + export PYTHONUNBUFFERED=1 + # Native crashes (SIGSEGV in a kernel) print every rank's Python stack. + export PYTHONFAULTHANDLER=1 + uv run --no-sync python dev/trainer_rank_landing_acceptance.py --phase contract + + cell() { # tp cp ep etp cell layers group + ART_MEGATRON_TENSOR_MODEL_PARALLEL_SIZE="$1" ART_MEGATRON_CONTEXT_PARALLEL_SIZE="$2" \ + ART_MEGATRON_EXPERT_MODEL_PARALLEL_SIZE="$3" ART_MEGATRON_EXPERT_TENSOR_PARALLEL_SIZE="$4" \ + timeout 7200s megatron_runtime/.venv/bin/python -m torch.distributed.run --standalone --nproc-per-node=$(( $1 * $2 )) \ + dev/trainer_rank_landing_acceptance.py --phase cost-calibrate \ + --cell "$5" --model "${MODEL}" --layers "$6" --group "$7" --repeat "${REPEAT}" \ + --evidence "${root}/evidence.jsonl" 2>&1 | tee "${root}/tp$1-cp$2-ep$3-etp$4-$5-$6-g$7.log" | { grep -E "cost-calibrate|LANDING|Traceback|outside|matches no calibrated" || true; } + } + # pipefail carries torchrun's status through tee and the grep filter; a + # failed cell is recorded and the campaign exits nonzero at the end. + failures=0 + run_cell() { cell "$@" || { failures=$((failures + 1)); echo "CELL FAILED: $*"; }; } + for shape in ${SHAPES}; do + IFS=, read -r tp cp ep etp <<< "${shape}" + etp="${etp:-1}" + for spec in ${CELLS}; do + IFS=: read -r name layers group <<< "${spec}" + run_cell "${tp}" "${cp}" "${ep}" "${etp}" "${name}" "${layers:-0}" "${group:-0}" + done + done + echo "cost-calibration lattice (${MODEL}): done (${failures} failed cells)" + [ "${failures}" -eq 0 ] + +config: + kubernetes: + pod_config: + spec: + schedulerName: binpack-scheduler + # A full five-shape dense lattice on a 27B-class model runs ~13 h of + # cells after setup; the 42000 s used by the smaller recipes cut it + # short and lost the run. The GPU pod policy caps this at 604800. + activeDeadlineSeconds: 86400 + containers: + - name: ray-node + imagePullPolicy: Always + env: + - name: UV_LINK_MODE + value: copy diff --git a/dev/trainer_rank_cost_calibration_manifest_attn-moe-h2048-h200-bf16.json b/dev/trainer_rank_cost_calibration_manifest_attn-moe-h2048-h200-bf16.json new file mode 100644 index 000000000..4a33cab20 --- /dev/null +++ b/dev/trainer_rank_cost_calibration_manifest_attn-moe-h2048-h200-bf16.json @@ -0,0 +1,131 @@ +{ + "campaign_labels": [ + + ], + "cells": [ + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g0","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g3","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g5","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g6","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g7","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":2,"ep":2,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g0","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":2,"ep":2,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":2,"ep":2,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":2,"ep":2,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g3","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":2,"ep":2,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":2,"ep":2,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g5","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":2,"ep":2,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g6","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":2,"ep":2,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2|g7","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g0","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g3","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g5","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g6","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|g7","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":2,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g0","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":2,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":2,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":2,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g3","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":2,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":2,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g5","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":2,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g6","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":2,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2|g7","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":4,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g0","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":4,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":4,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":4,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g3","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":4,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":4,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g5","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":4,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g6","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":4,"ep":4,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4|g7","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g0","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g3","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g5","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g6","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|g7","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":1,"ep":2,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g0","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":1,"ep":2,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":1,"ep":2,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":1,"ep":2,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g3","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":1,"ep":2,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":1,"ep":2,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g5","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":1,"ep":2,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g6","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-ellavox-qwen3","cp":1,"ep":2,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2|g7","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g0","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g3","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g5","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g6","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|g7","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-grpo-g16","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-grpo-g16","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-grpo-g16","cp":2,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-grpo-g16","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-grpo-g16","cp":4,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-grpo-g16","cp":4,"ep":4,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-grpo-g16","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-grpo-g16","cp":1,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-grpo-g4x4","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-grpo-g4x4","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-grpo-g4x4","cp":2,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-grpo-g4x4","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-grpo-g4x4","cp":4,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-grpo-g4x4","cp":4,"ep":4,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-grpo-g4x4","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-grpo-g4x4","cp":1,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-grpo-g8","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-grpo-g8","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-grpo-g8","cp":2,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-grpo-g8","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-grpo-g8","cp":4,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-grpo-g8","cp":4,"ep":4,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-grpo-g8","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-grpo-g8","cp":1,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-hetero2","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-hetero2","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-hetero2","cp":2,"ep":2,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-hetero2","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-hetero2","cp":4,"ep":2,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-hetero2","cp":4,"ep":4,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-hetero2","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-hetero2","cp":1,"ep":2,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-hetero3","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-hetero3","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-hetero3","cp":2,"ep":2,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-hetero3","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-hetero3","cp":4,"ep":2,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-hetero3","cp":4,"ep":4,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-hetero3","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-hetero3","cp":1,"ep":2,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-hetero","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-hetero","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-hetero","cp":2,"ep":2,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp2|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-hetero","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-hetero","cp":4,"ep":2,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-hetero","cp":4,"ep":4,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp1|cp4|ep4","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":1}, + {"campaign":"tr-cost-q3-30b-ep1","cell":"cal-hetero","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp2|cp1","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2}, + {"campaign":"tr-cost-q3-30b-ep","cell":"cal-hetero","cp":1,"ep":2,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-30B-A3B|L48|tp2|cp1|ep2","layers":48,"model":"Qwen/Qwen3-30B-A3B","tp":2} + ], + "description": "Expected calibration cells for the Qwen3-30B-A3B (attention + MoE, hidden 2,048, 128 experts top-8) execution class on H200 bf16 (2026-09-04): the exact cells of the two shape-lattice launches below (dev/trainer_rank_cost_calibration_lattice.sky.yaml; EP1 at TP1 x CP1/2/4 and TP2 x CP1, EP2 at CP2/CP4/TP2, EP4 at CP4) on the synthetic families and the Qwen3-tokenized Ellavox groups, keyed exactly as the fitter keys cells. TP1 x CP2 (EP1/EP2) and TP1 x CP4 (EP1/EP4) were re-measured with sixteen rounds (launch tr-cost-q3-30b-r16) after the eight-round timings carried sporadic multi-second stalls; those shapes' cells come from that launch. Admitted directly at TP1 x CP1, TP1 x CP2 (EP1, EP2), TP2 x CP1 (EP1, EP2); TP1 x CP4 (EP1, EP2, EP4) through the two-stage re-ranker.", + "excluded": [ + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g6","reason":"every candidate refused by memory admission on one H200 (issue #848); no measurable rows"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-30B-A3B|L48|tp1|cp1|g7","reason":"cold memory admission lets the 94,721-token group run on one H200 and it hits CUDA out-of-memory (issue #848); no measurable rows"}, + {"key":"cal-grpo-g16|Qwen/Qwen3-30B-A3B|L48|tp1|cp1","reason":"cold memory admission lets the 106k-token forward run on one H200 and it hits CUDA out-of-memory (issue #848); no measurable rows"} + ], + "launches": {"tr-cost-q3-30b-ep":{"cells":["cal-grpo-g8:0","cal-grpo-g16:0","cal-grpo-g4x4:0","cal-hetero:0","cal-hetero2:0","cal-hetero3:0","cal-ellavox-qwen3:0:0","cal-ellavox-qwen3:0:1","cal-ellavox-qwen3:0:2","cal-ellavox-qwen3:0:3","cal-ellavox-qwen3:0:4","cal-ellavox-qwen3:0:5","cal-ellavox-qwen3:0:6","cal-ellavox-qwen3:0:7"],"model":"Qwen/Qwen3-30B-A3B","recipe":"dev/trainer_rank_cost_calibration_lattice.sky.yaml","repeat":8,"shapes":[[1,2,2,1],[1,4,2,1],[1,4,4,1],[2,1,2,1]]},"tr-cost-q3-30b-ep1":{"cells":["cal-grpo-g8:0","cal-grpo-g16:0","cal-grpo-g4x4:0","cal-hetero:0","cal-hetero2:0","cal-hetero3:0","cal-ellavox-qwen3:0:0","cal-ellavox-qwen3:0:1","cal-ellavox-qwen3:0:2","cal-ellavox-qwen3:0:3","cal-ellavox-qwen3:0:4","cal-ellavox-qwen3:0:5","cal-ellavox-qwen3:0:6","cal-ellavox-qwen3:0:7"],"model":"Qwen/Qwen3-30B-A3B","recipe":"dev/trainer_rank_cost_calibration_lattice.sky.yaml","repeat":8,"shapes":[[1,1,1,1],[1,2,1,1],[1,4,1,1],[2,1,1,1]]},"tr-cost-q3-30b-r16":{"cells":["cal-grpo-g8:0","cal-grpo-g16:0","cal-grpo-g4x4:0","cal-hetero:0","cal-hetero2:0","cal-hetero3:0","cal-ellavox-qwen3:0:0","cal-ellavox-qwen3:0:1","cal-ellavox-qwen3:0:2","cal-ellavox-qwen3:0:3","cal-ellavox-qwen3:0:4","cal-ellavox-qwen3:0:5","cal-ellavox-qwen3:0:6","cal-ellavox-qwen3:0:7"],"model":"Qwen/Qwen3-30B-A3B","recipe":"dev/trainer_rank_cost_calibration_lattice.sky.yaml","repeat":16,"shapes":[[1,2,1,1],[1,4,1,1],[1,2,2,1],[1,4,4,1]]}}, + "recipes": [ + + ], + "schema": "art.dev.trainer_rank_cost_calibration_manifest.v1", + "table_id": "attn-moe-h2048-h200-bf16" +} diff --git a/dev/trainer_rank_cost_calibration_manifest_dense-attn-h2048-h200-bf16.json b/dev/trainer_rank_cost_calibration_manifest_dense-attn-h2048-h200-bf16.json new file mode 100644 index 000000000..33ee868f6 --- /dev/null +++ b/dev/trainer_rank_cost_calibration_manifest_dense-attn-h2048-h200-bf16.json @@ -0,0 +1,100 @@ +{ + "campaign_labels": [ + + ], + "cells": [ + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g0","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g3","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g4","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g5","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g6","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp1|g7","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g0","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g3","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g4","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g5","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g6","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp2|g7","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g0","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g3","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g4","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g5","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g6","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp1|cp4|g7","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g0","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g3","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g4","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g5","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g6","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp1|g7","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g0","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g3","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g4","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g5","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g6","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g7","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g16","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp1|cp1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g16","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp1|cp2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g16","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp1|cp4","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g16","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp2|cp1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g16","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp2|cp2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g4x4","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp1|cp1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g4x4","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp1|cp2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g4x4","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp1|cp4","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g4x4","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp2|cp1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g4x4","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp2|cp2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g8","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp1|cp1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g8","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp1|cp2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g8","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp1|cp4","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g8","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp2|cp1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-grpo-g8","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp2|cp2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero2","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp1|cp1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero2","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp1|cp2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero2","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp1|cp4","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero2","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp2|cp1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero2","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp2|cp2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero3","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-1.7B|L28|tp1|cp1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero3","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-1.7B|L28|tp1|cp2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero3","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-1.7B|L28|tp1|cp4","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero3","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-1.7B|L28|tp2|cp1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero3","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-1.7B|L28|tp2|cp2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-1.7B|L28|tp1|cp1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-1.7B|L28|tp1|cp2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-1.7B|L28|tp1|cp4","layers":28,"model":"Qwen/Qwen3-1.7B","tp":1}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-1.7B|L28|tp2|cp1","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2}, + {"campaign":"tr-cost-q3-1p7b","cell":"cal-hetero","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-1.7B|L28|tp2|cp2","layers":28,"model":"Qwen/Qwen3-1.7B","tp":2} + ], + "description": "Expected calibration cells for the Qwen3-1.7B (dense attention, hidden 2,048) execution class on H200 bf16 (2026-09-04): the exact cells of the shape-lattice launch below (dev/trainer_rank_cost_calibration_lattice.sky.yaml; TP1 x CP1/2/4, TP2 x CP1 and TP2 x CP2), keyed exactly as the fitter keys cells. TP1 x CP4 is selected by the two-stage re-ranker (the ten-term score shortlists, the CP plan structure decides), certified in the certificate's reranker block. TP2 x CP2 is measured but excluded from the table (fails its gates).", + "excluded": [ + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g0","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g1","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g3","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g4","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g5","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g6","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-1.7B|L28|tp2|cp2|g7","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-grpo-g16|Qwen/Qwen3-1.7B|L28|tp2|cp2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-grpo-g4x4|Qwen/Qwen3-1.7B|L28|tp2|cp2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-grpo-g8|Qwen/Qwen3-1.7B|L28|tp2|cp2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-hetero2|Qwen/Qwen3-1.7B|L28|tp2|cp2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-hetero3|Qwen/Qwen3-1.7B|L28|tp2|cp2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-hetero|Qwen/Qwen3-1.7B|L28|tp2|cp2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"} + ], + "launches": {"tr-cost-q3-1p7b":{"cells":["cal-grpo-g8:0","cal-grpo-g16:0","cal-grpo-g4x4:0","cal-hetero:0","cal-hetero2:0","cal-hetero3:0","cal-ellavox-qwen3:0:0","cal-ellavox-qwen3:0:1","cal-ellavox-qwen3:0:2","cal-ellavox-qwen3:0:3","cal-ellavox-qwen3:0:4","cal-ellavox-qwen3:0:5","cal-ellavox-qwen3:0:6","cal-ellavox-qwen3:0:7"],"model":"Qwen/Qwen3-1.7B","recipe":"dev/trainer_rank_cost_calibration_lattice.sky.yaml","repeat":8,"shapes":[[1,1,1,1],[1,2,1,1],[1,4,1,1],[2,1,1,1],[2,2,1,1]]}}, + "recipes": [ + + ], + "schema": "art.dev.trainer_rank_cost_calibration_manifest.v1", + "table_id": "dense-attn-h2048-h200-bf16" +} \ No newline at end of file diff --git a/dev/trainer_rank_cost_calibration_manifest_dense-attn-h4096-h200-bf16.json b/dev/trainer_rank_cost_calibration_manifest_dense-attn-h4096-h200-bf16.json new file mode 100644 index 000000000..2709f3364 --- /dev/null +++ b/dev/trainer_rank_cost_calibration_manifest_dense-attn-h4096-h200-bf16.json @@ -0,0 +1,100 @@ +{ + "campaign_labels": [ + + ], + "cells": [ + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g0","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g1","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g2","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g3","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g4","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g5","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g6","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp1|g7","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g0","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g1","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g2","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g3","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g4","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g5","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g6","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp2|g7","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g0","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g1","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g2","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g3","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g4","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g5","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g6","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp1|cp4|g7","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g0","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g1","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g2","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g3","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g4","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g5","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g6","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp1|g7","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g0","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g1","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g2","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g3","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g4","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g5","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g6","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g7","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g16","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-8B|L36|tp1|cp1","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g16","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-8B|L36|tp1|cp2","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g16","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-8B|L36|tp1|cp4","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g16","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-8B|L36|tp2|cp1","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g16","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-8B|L36|tp2|cp2","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g4x4","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp1|cp1","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g4x4","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp1|cp2","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g4x4","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp1|cp4","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g4x4","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp2|cp1","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g4x4","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp2|cp2","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g8","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-8B|L36|tp1|cp1","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g8","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-8B|L36|tp1|cp2","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g8","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-8B|L36|tp1|cp4","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g8","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-8B|L36|tp2|cp1","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-grpo-g8","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-8B|L36|tp2|cp2","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero2","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-8B|L36|tp1|cp1","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero2","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-8B|L36|tp1|cp2","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero2","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-8B|L36|tp1|cp4","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero2","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-8B|L36|tp2|cp1","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero2","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-8B|L36|tp2|cp2","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero3","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-8B|L36|tp1|cp1","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero3","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-8B|L36|tp1|cp2","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero3","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-8B|L36|tp1|cp4","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero3","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-8B|L36|tp2|cp1","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero3","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-8B|L36|tp2|cp2","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-8B|L36|tp1|cp1","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-8B|L36|tp1|cp2","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-8B|L36|tp1|cp4","layers":36,"model":"Qwen/Qwen3-8B","tp":1}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-8B|L36|tp2|cp1","layers":36,"model":"Qwen/Qwen3-8B","tp":2}, + {"campaign":"tr-cost-q3-8b","cell":"cal-hetero","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-8B|L36|tp2|cp2","layers":36,"model":"Qwen/Qwen3-8B","tp":2} + ], + "description": "Expected calibration cells for the Qwen3-8B (dense attention, hidden 4,096) execution class on H200 bf16 (2026-09-04): the exact cells of the shape-lattice launch below (dev/trainer_rank_cost_calibration_lattice.sky.yaml; TP1 x CP1/2/4, TP2 x CP1 and TP2 x CP2), keyed exactly as the fitter keys cells. TP1 x CP4 is selected by the two-stage re-ranker (the ten-term score shortlists, the CP plan structure decides), certified in the certificate's reranker block. TP2 x CP2 is measured but excluded from the table (fails its gates).", + "excluded": [ + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g0","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g1","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g3","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g4","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g5","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g6","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-8B|L36|tp2|cp2|g7","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-grpo-g16|Qwen/Qwen3-8B|L36|tp2|cp2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-grpo-g4x4|Qwen/Qwen3-8B|L36|tp2|cp2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-grpo-g8|Qwen/Qwen3-8B|L36|tp2|cp2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-hetero2|Qwen/Qwen3-8B|L36|tp2|cp2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-hetero3|Qwen/Qwen3-8B|L36|tp2|cp2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"}, + {"key":"cal-hetero|Qwen/Qwen3-8B|L36|tp2|cp2","reason":"TP2 x CP2 fails the per-shape gates for this class (the CP exchange-schedule effect; a re-ranker for it is not certified); not admitted"} + ], + "launches": {"tr-cost-q3-8b":{"cells":["cal-grpo-g8:0","cal-grpo-g16:0","cal-grpo-g4x4:0","cal-hetero:0","cal-hetero2:0","cal-hetero3:0","cal-ellavox-qwen3:0:0","cal-ellavox-qwen3:0:1","cal-ellavox-qwen3:0:2","cal-ellavox-qwen3:0:3","cal-ellavox-qwen3:0:4","cal-ellavox-qwen3:0:5","cal-ellavox-qwen3:0:6","cal-ellavox-qwen3:0:7"],"model":"Qwen/Qwen3-8B","recipe":"dev/trainer_rank_cost_calibration_lattice.sky.yaml","repeat":8,"shapes":[[1,1,1,1],[1,2,1,1],[1,4,1,1],[2,1,1,1],[2,2,1,1]]}}, + "recipes": [ + + ], + "schema": "art.dev.trainer_rank_cost_calibration_manifest.v1", + "table_id": "dense-attn-h4096-h200-bf16" +} \ No newline at end of file diff --git a/dev/trainer_rank_cost_calibration_manifest_dense-attn-h5120-h200-bf16.json b/dev/trainer_rank_cost_calibration_manifest_dense-attn-h5120-h200-bf16.json new file mode 100644 index 000000000..57c4e497f --- /dev/null +++ b/dev/trainer_rank_cost_calibration_manifest_dense-attn-h5120-h200-bf16.json @@ -0,0 +1,87 @@ +{ + "campaign_labels": [ + + ], + "cells": [ + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g0","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g1","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g2","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g3","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g4","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g5","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g6","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp1|g7","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g0","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g1","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g2","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g3","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g4","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g5","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g6","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp2|g7","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g0","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g1","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g2","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g3","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g4","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g5","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g6","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp1|cp4|g7","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g0","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g1","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g2","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g3","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g4","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g5","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g6","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":1,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp1|g7","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g0","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g1","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g2","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g3","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g4","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g5","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g6","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-14B|L40|tp2|cp2|g7","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g16","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-14B|L40|tp1|cp1","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g16","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-14B|L40|tp1|cp2","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g16","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-14B|L40|tp1|cp4","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g16","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-14B|L40|tp2|cp1","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g16","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3-14B|L40|tp2|cp2","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g4x4","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp1|cp1","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g4x4","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp1|cp2","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g4x4","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp1|cp4","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g4x4","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp2|cp1","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g4x4","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3-14B|L40|tp2|cp2","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g8","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-14B|L40|tp1|cp1","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g8","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-14B|L40|tp1|cp2","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g8","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-14B|L40|tp1|cp4","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g8","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-14B|L40|tp2|cp1","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-grpo-g8","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3-14B|L40|tp2|cp2","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero2","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-14B|L40|tp1|cp1","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero2","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-14B|L40|tp1|cp2","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero2","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-14B|L40|tp1|cp4","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero2","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-14B|L40|tp2|cp1","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero2","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3-14B|L40|tp2|cp2","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero3","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-14B|L40|tp1|cp1","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero3","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-14B|L40|tp1|cp2","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero3","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-14B|L40|tp1|cp4","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero3","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-14B|L40|tp2|cp1","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero3","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3-14B|L40|tp2|cp2","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-14B|L40|tp1|cp1","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-14B|L40|tp1|cp2","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-14B|L40|tp1|cp4","layers":40,"model":"Qwen/Qwen3-14B","tp":1}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-14B|L40|tp2|cp1","layers":40,"model":"Qwen/Qwen3-14B","tp":2}, + {"campaign":"tr-cost-q3-14b","cell":"cal-hetero","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3-14B|L40|tp2|cp2","layers":40,"model":"Qwen/Qwen3-14B","tp":2} + ], + "description": "Expected calibration cells for the Qwen3-14B (dense attention, hidden 5,120) execution class on H200 bf16 (2026-09-04): the exact cells of the shape-lattice launch below (dev/trainer_rank_cost_calibration_lattice.sky.yaml; TP1 x CP1/2/4, TP2 x CP1 and TP2 x CP2), keyed exactly as the fitter keys cells. TP1 x CP4 is selected by the two-stage re-ranker (the ten-term score shortlists, the CP plan structure decides), certified in the certificate's reranker block.", + "excluded": [ + + ], + "launches": {"tr-cost-q3-14b":{"cells":["cal-grpo-g8:0","cal-grpo-g16:0","cal-grpo-g4x4:0","cal-hetero:0","cal-hetero2:0","cal-hetero3:0","cal-ellavox-qwen3:0:0","cal-ellavox-qwen3:0:1","cal-ellavox-qwen3:0:2","cal-ellavox-qwen3:0:3","cal-ellavox-qwen3:0:4","cal-ellavox-qwen3:0:5","cal-ellavox-qwen3:0:6","cal-ellavox-qwen3:0:7"],"model":"Qwen/Qwen3-14B","recipe":"dev/trainer_rank_cost_calibration_lattice.sky.yaml","repeat":8,"shapes":[[1,1,1,1],[1,2,1,1],[1,4,1,1],[2,1,1,1],[2,2,1,1]]}}, + "recipes": [ + + ], + "schema": "art.dev.trainer_rank_cost_calibration_manifest.v1", + "table_id": "dense-attn-h5120-h200-bf16" +} \ No newline at end of file diff --git a/dev/trainer_rank_cost_calibration_manifest_dense-gdn-h5120-h200-bf16.json b/dev/trainer_rank_cost_calibration_manifest_dense-gdn-h5120-h200-bf16.json new file mode 100644 index 000000000..7932c582d --- /dev/null +++ b/dev/trainer_rank_cost_calibration_manifest_dense-gdn-h5120-h200-bf16.json @@ -0,0 +1,104 @@ +{ + "campaign_labels": [ + + ], + "cells": [ + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":0,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g0","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":1,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":2,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":3,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g3","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":4,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g4","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":5,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g5","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":6,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g6","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":7,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g7","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":0,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g0","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":1,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":2,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":3,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g3","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":4,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g4","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":5,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g5","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":6,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g6","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":7,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g7","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":0,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g0","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":1,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":2,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":3,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g3","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":4,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g4","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":5,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g5","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":6,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g6","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":7,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp4|g7","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":0,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g0","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":1,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":2,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":3,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g3","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":4,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g4","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":5,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g5","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":6,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g6","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":7,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp1|g7","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":0,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g0","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":1,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":2,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":3,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g3","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":4,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g4","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":5,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g5","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":6,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g6","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":7,"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g7","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-grpo-g16","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp1|cp1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-grpo-g16","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp1|cp2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-grpo-g16","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp1|cp4","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-grpo-g16","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp2|cp1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-grpo-g16","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp2|cp2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-grpo-g4x4","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp1|cp1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-grpo-g4x4","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp1|cp2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-grpo-g4x4","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp1|cp4","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-grpo-g4x4","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp2|cp1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-grpo-g4x4","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp2|cp2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-grpo-g8","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp1|cp1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-grpo-g8","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp1|cp2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-grpo-g8","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp1|cp4","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-grpo-g8","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp2|cp1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-grpo-g8","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp2|cp2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-hetero2","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3.5-27B|L64|tp1|cp1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-hetero2","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3.5-27B|L64|tp1|cp2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-hetero2","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3.5-27B|L64|tp1|cp4","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-hetero2","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3.5-27B|L64|tp2|cp1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-hetero2","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3.5-27B|L64|tp2|cp2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-hetero3","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3.5-27B|L64|tp1|cp1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-hetero3","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3.5-27B|L64|tp1|cp2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-hetero3","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3.5-27B|L64|tp1|cp4","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-hetero3","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3.5-27B|L64|tp2|cp1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-hetero3","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3.5-27B|L64|tp2|cp2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-hetero","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-27B|L64|tp1|cp1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-a","cell":"cal-hetero","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-27B|L64|tp1|cp2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-hetero","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-27B|L64|tp1|cp4","layers":64,"model":"Qwen/Qwen3.5-27B","tp":1}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-hetero","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-27B|L64|tp2|cp1","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2}, + {"campaign":"tr-cost-q35-27b-b","cell":"cal-hetero","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-27B|L64|tp2|cp2","layers":64,"model":"Qwen/Qwen3.5-27B","tp":2} + ], + "description": "Expected calibration cells for the Qwen3.5-27B (dense GDN + attention, hidden 5,120) execution class on H200 bf16 (2026-09-04): the exact cells of the two shape-lattice launches below (dev/trainer_rank_cost_calibration_lattice.sky.yaml; TP1 x CP1/2/4, TP2 x CP1, TP2 x CP2) on the synthetic families and the Qwen3.5-tokenized Ellavox groups, keyed exactly as the fitter keys cells. Admitted at TP1 x CP1/2/4 and TP2 x CP1; TP2 x CP2 is measured but excluded from the table (fails its gates).", + "excluded": [ + {"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g6","reason":"the unshared layout is refused by single-GPU memory admission; incomplete cell (the fitter never drops candidates silently)"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp1|g7","reason":"the unshared and two shallow layouts of the 94,721-token group are refused by single-GPU memory admission; incomplete cell"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp1|cp2|g7","reason":"the unshared and one shallow layout of the 94,721-token group are refused by memory admission at CP2; incomplete cell"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g0","reason":"TP2 x CP2 fails its gates for this class (p95 regret 5.12% on the Ellavox g6 cell, a clear winner selected >5% off); not admitted"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g1","reason":"TP2 x CP2 fails its gates for this class (p95 regret 5.12% on the Ellavox g6 cell, a clear winner selected >5% off); not admitted"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g2","reason":"TP2 x CP2 fails its gates for this class (p95 regret 5.12% on the Ellavox g6 cell, a clear winner selected >5% off); not admitted"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g3","reason":"TP2 x CP2 fails its gates for this class (p95 regret 5.12% on the Ellavox g6 cell, a clear winner selected >5% off); not admitted"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g4","reason":"TP2 x CP2 fails its gates for this class (p95 regret 5.12% on the Ellavox g6 cell, a clear winner selected >5% off); not admitted"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g5","reason":"TP2 x CP2 fails its gates for this class (p95 regret 5.12% on the Ellavox g6 cell, a clear winner selected >5% off); not admitted"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g6","reason":"TP2 x CP2 fails its gates for this class (p95 regret 5.12% on the Ellavox g6 cell, a clear winner selected >5% off); not admitted"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-27B|L64|tp2|cp2|g7","reason":"TP2 x CP2 fails its gates for this class (p95 regret 5.12% on the Ellavox g6 cell, a clear winner selected >5% off); not admitted"}, + {"key":"cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp1|cp1","reason":"cold memory admission lets the 106k-token forward run on one H200 and it hits CUDA out-of-memory (issue #848); no measurable rows"}, + {"key":"cal-grpo-g16|Qwen/Qwen3.5-27B|L64|tp2|cp2","reason":"TP2 x CP2 fails its gates for this class (p95 regret 5.12% on the Ellavox g6 cell, a clear winner selected >5% off); not admitted"}, + {"key":"cal-grpo-g4x4|Qwen/Qwen3.5-27B|L64|tp2|cp2","reason":"TP2 x CP2 fails its gates for this class (p95 regret 5.12% on the Ellavox g6 cell, a clear winner selected >5% off); not admitted"}, + {"key":"cal-grpo-g8|Qwen/Qwen3.5-27B|L64|tp2|cp2","reason":"TP2 x CP2 fails its gates for this class (p95 regret 5.12% on the Ellavox g6 cell, a clear winner selected >5% off); not admitted"}, + {"key":"cal-hetero2|Qwen/Qwen3.5-27B|L64|tp2|cp2","reason":"NCCL collective timeout in the tensor-parallel group during warm-up (two ranks aborted); no measurable rows"}, + {"key":"cal-hetero3|Qwen/Qwen3.5-27B|L64|tp2|cp2","reason":"TP2 x CP2 fails its gates for this class (p95 regret 5.12% on the Ellavox g6 cell, a clear winner selected >5% off); not admitted"}, + {"key":"cal-hetero|Qwen/Qwen3.5-27B|L64|tp2|cp2","reason":"TP2 x CP2 fails its gates for this class (p95 regret 5.12% on the Ellavox g6 cell, a clear winner selected >5% off); not admitted"} + ], + "launches": {"tr-cost-q35-27b-a":{"cells":["cal-grpo-g8:0","cal-grpo-g16:0","cal-grpo-g4x4:0","cal-hetero:0","cal-hetero2:0","cal-hetero3:0","cal-ellavox:0:0","cal-ellavox:0:1","cal-ellavox:0:2","cal-ellavox:0:3","cal-ellavox:0:4","cal-ellavox:0:5","cal-ellavox:0:6","cal-ellavox:0:7"],"model":"Qwen/Qwen3.5-27B","recipe":"dev/trainer_rank_cost_calibration_lattice.sky.yaml","repeat":8,"shapes":[[1,1,1,1],[1,2,1,1]]},"tr-cost-q35-27b-b":{"cells":["cal-grpo-g8:0","cal-grpo-g16:0","cal-grpo-g4x4:0","cal-hetero:0","cal-hetero2:0","cal-hetero3:0","cal-ellavox:0:0","cal-ellavox:0:1","cal-ellavox:0:2","cal-ellavox:0:3","cal-ellavox:0:4","cal-ellavox:0:5","cal-ellavox:0:6","cal-ellavox:0:7"],"model":"Qwen/Qwen3.5-27B","recipe":"dev/trainer_rank_cost_calibration_lattice.sky.yaml","repeat":8,"shapes":[[1,4,1,1],[2,1,1,1],[2,2,1,1]]}}, + "recipes": [ + + ], + "schema": "art.dev.trainer_rank_cost_calibration_manifest.v1", + "table_id": "dense-gdn-h5120-h200-bf16" +} diff --git a/dev/trainer_rank_cost_calibration_manifest.json b/dev/trainer_rank_cost_calibration_manifest_dense-h2560-h200-bf16.json similarity index 52% rename from dev/trainer_rank_cost_calibration_manifest.json rename to dev/trainer_rank_cost_calibration_manifest_dense-h2560-h200-bf16.json index 0337ce0eb..a1e9df5ab 100644 --- a/dev/trainer_rank_cost_calibration_manifest.json +++ b/dev/trainer_rank_cost_calibration_manifest_dense-h2560-h200-bf16.json @@ -1,4 +1,12 @@ { + "campaign_labels": [ + "local", + "tr-cost-2gpu", + "tr-cost-2gpu-2", + "tr-cost-cp4", + "tr-cost-cp4-2", + "tr-cost-cp4-840" + ], "cells": [ {"campaign":"local","cell":"cal-ellavox","cp":1,"group":0,"key":"cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g0","layers":32,"model":"Qwen/Qwen3.5-4B","tp":1}, {"campaign":"local","cell":"cal-ellavox","cp":1,"group":1,"key":"cal-ellavox|Qwen/Qwen3.5-4B|L32|tp1|cp1|g1","layers":32,"model":"Qwen/Qwen3.5-4B","tp":1}, @@ -57,11 +65,52 @@ {"campaign":"local","cell":"cal-hetero","cp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-4B|L32|tp1|cp1","layers":32,"model":"Qwen/Qwen3.5-4B","tp":1}, {"campaign":"tr-cost-2gpu-2","cell":"cal-hetero","cp":2,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-4B|L32|tp1|cp2","layers":32,"model":"Qwen/Qwen3.5-4B","tp":1}, {"campaign":"tr-cost-cp4","cell":"cal-hetero","cp":4,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-4B|L32|tp1|cp4","layers":32,"model":"Qwen/Qwen3.5-4B","tp":1}, - {"campaign":"tr-cost-2gpu","cell":"cal-hetero","cp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-4B|L32|tp2|cp1","layers":32,"model":"Qwen/Qwen3.5-4B","tp":2} + {"campaign":"tr-cost-2gpu","cell":"cal-hetero","cp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-4B|L32|tp2|cp1","layers":32,"model":"Qwen/Qwen3.5-4B","tp":2}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g0","layers":36,"model":"Qwen/Qwen3-4B","tp":1}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g1","layers":36,"model":"Qwen/Qwen3-4B","tp":1}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g2","layers":36,"model":"Qwen/Qwen3-4B","tp":1}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g3","layers":36,"model":"Qwen/Qwen3-4B","tp":1}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g4","layers":36,"model":"Qwen/Qwen3-4B","tp":1}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g5","layers":36,"model":"Qwen/Qwen3-4B","tp":1}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g6","layers":36,"model":"Qwen/Qwen3-4B","tp":1}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":4,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g7","layers":36,"model":"Qwen/Qwen3-4B","tp":1}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":0,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g0","layers":36,"model":"Qwen/Qwen3-4B","tp":2}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":1,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g1","layers":36,"model":"Qwen/Qwen3-4B","tp":2}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":2,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g2","layers":36,"model":"Qwen/Qwen3-4B","tp":2}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":3,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g3","layers":36,"model":"Qwen/Qwen3-4B","tp":2}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":4,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g4","layers":36,"model":"Qwen/Qwen3-4B","tp":2}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":5,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g5","layers":36,"model":"Qwen/Qwen3-4B","tp":2}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":6,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g6","layers":36,"model":"Qwen/Qwen3-4B","tp":2}, + {"campaign":"tr-cost-q3-4b-cp4","cell":"cal-ellavox-qwen3","cp":2,"ep":1,"etp":1,"group":7,"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g7","layers":36,"model":"Qwen/Qwen3-4B","tp":2} ], "description": "Expected calibration cells for the coefficient-version-2 campaign (2026-09-02; Ellavox CP4 groups 1 and 4 re-measured 2026-09-03 after the harness desynchronization fix of issue #840): the exact union of the cells the checked-in recipes launch (dev/trainer_rank_cost_calibration_local.sh; the cp4 and 2gpu sky recipes in both CELL_SET modes), keyed exactly as the fitter keys cells. tests/unit/test_planner_cost_manifest_recipes.py asserts recipes and manifest agree. The fitter's --manifest validation requires every non-excluded cell to be present with a complete mandatory candidate set and rejects unexpected cells; exclusions must be listed here with a reason.", "excluded": [ - + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g0","reason":"blind-spot evidence, not fitted: Qwen3-4B at TP1 x CP4 on the Qwen3-tokenized Ellavox groups shows the shipped score misranking layouts (regret up to 35% on g5) because the ten terms cannot see the context-parallel exchange schedule; the pair is withheld from admission (see withheld)"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g1","reason":"blind-spot evidence, not fitted: Qwen3-4B at TP1 x CP4 on the Qwen3-tokenized Ellavox groups shows the shipped score misranking layouts (regret up to 35% on g5) because the ten terms cannot see the context-parallel exchange schedule; the pair is withheld from admission (see withheld)"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g2","reason":"blind-spot evidence, not fitted: Qwen3-4B at TP1 x CP4 on the Qwen3-tokenized Ellavox groups shows the shipped score misranking layouts (regret up to 35% on g5) because the ten terms cannot see the context-parallel exchange schedule; the pair is withheld from admission (see withheld)"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g3","reason":"blind-spot evidence, not fitted: Qwen3-4B at TP1 x CP4 on the Qwen3-tokenized Ellavox groups shows the shipped score misranking layouts (regret up to 35% on g5) because the ten terms cannot see the context-parallel exchange schedule; the pair is withheld from admission (see withheld)"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g4","reason":"blind-spot evidence, not fitted: Qwen3-4B at TP1 x CP4 on the Qwen3-tokenized Ellavox groups shows the shipped score misranking layouts (regret up to 35% on g5) because the ten terms cannot see the context-parallel exchange schedule; the pair is withheld from admission (see withheld)"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g5","reason":"blind-spot evidence, not fitted: Qwen3-4B at TP1 x CP4 on the Qwen3-tokenized Ellavox groups shows the shipped score misranking layouts (regret up to 35% on g5) because the ten terms cannot see the context-parallel exchange schedule; the pair is withheld from admission (see withheld)"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g6","reason":"blind-spot evidence, not fitted: Qwen3-4B at TP1 x CP4 on the Qwen3-tokenized Ellavox groups shows the shipped score misranking layouts (regret up to 35% on g5) because the ten terms cannot see the context-parallel exchange schedule; the pair is withheld from admission (see withheld)"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp1|cp4|g7","reason":"blind-spot evidence, not fitted: Qwen3-4B at TP1 x CP4 on the Qwen3-tokenized Ellavox groups shows the shipped score misranking layouts (regret up to 35% on g5) because the ten terms cannot see the context-parallel exchange schedule; the pair is withheld from admission (see withheld)"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g0","reason":"blind-spot evidence, not fitted: TP2 x CP2 is not an admitted shape of this table; measured to show the same context-parallel exchange-schedule effect for the attention geometry"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g1","reason":"blind-spot evidence, not fitted: TP2 x CP2 is not an admitted shape of this table; measured to show the same context-parallel exchange-schedule effect for the attention geometry"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g2","reason":"blind-spot evidence, not fitted: TP2 x CP2 is not an admitted shape of this table; measured to show the same context-parallel exchange-schedule effect for the attention geometry"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g3","reason":"blind-spot evidence, not fitted: TP2 x CP2 is not an admitted shape of this table; measured to show the same context-parallel exchange-schedule effect for the attention geometry"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g4","reason":"blind-spot evidence, not fitted: TP2 x CP2 is not an admitted shape of this table; measured to show the same context-parallel exchange-schedule effect for the attention geometry"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g5","reason":"blind-spot evidence, not fitted: TP2 x CP2 is not an admitted shape of this table; measured to show the same context-parallel exchange-schedule effect for the attention geometry"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g6","reason":"blind-spot evidence, not fitted: TP2 x CP2 is not an admitted shape of this table; measured to show the same context-parallel exchange-schedule effect for the attention geometry"}, + {"key":"cal-ellavox-qwen3|Qwen/Qwen3-4B|L36|tp2|cp2|g7","reason":"blind-spot evidence, not fitted: TP2 x CP2 is not an admitted shape of this table; measured to show the same context-parallel exchange-schedule effect for the attention geometry"} + ], + "launches": {"tr-cost-q3-4b-cp4":{"cells":["cal-ellavox-qwen3:0:0","cal-ellavox-qwen3:0:1","cal-ellavox-qwen3:0:2","cal-ellavox-qwen3:0:3","cal-ellavox-qwen3:0:4","cal-ellavox-qwen3:0:5","cal-ellavox-qwen3:0:6","cal-ellavox-qwen3:0:7"],"model":"Qwen/Qwen3-4B","recipe":"dev/trainer_rank_cost_calibration_lattice.sky.yaml","repeat":8,"shapes":[[1,4,1,1],[2,2,1,1]]}}, + "recipes": [ + "dev/trainer_rank_cost_calibration_local.sh", + "dev/trainer_rank_cost_calibration_cp4.sky.yaml", + "dev/trainer_rank_cost_calibration_2gpu.sky.yaml" ], - "schema": "art.dev.trainer_rank_cost_calibration_manifest.v1" -} + "schema": "art.dev.trainer_rank_cost_calibration_manifest.v1", + "table_id": "dense-h2560-h200-bf16", + "withheld": [ + {"model":"Qwen/Qwen3-4B","reason":"measured and fitted (three synthetic cells) but not admitted: on real-data groups the ten-term score cannot see the context-parallel exchange schedule and its regret reached 35% (dev/trainer_rank_planner_design.md, known limitation); version 1 is the less harmful fallback there until the CP-aware re-ranker is certified","shape":[1,4,1,1]} + ] +} \ No newline at end of file diff --git a/dev/trainer_rank_cost_calibration_manifest_gdn-moe-h2048-h200-bf16.json b/dev/trainer_rank_cost_calibration_manifest_gdn-moe-h2048-h200-bf16.json new file mode 100644 index 000000000..f03dca05c --- /dev/null +++ b/dev/trainer_rank_cost_calibration_manifest_gdn-moe-h2048-h200-bf16.json @@ -0,0 +1,119 @@ +{ + "campaign_labels": [ + + ], + "cells": [ + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":0,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g0","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":1,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":2,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":3,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g3","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":4,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":5,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g5","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":6,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g6","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":1,"ep":1,"etp":1,"group":7,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g7","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":2,"ep":2,"etp":1,"group":0,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g0","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":2,"ep":2,"etp":1,"group":1,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":2,"ep":2,"etp":1,"group":2,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":2,"ep":2,"etp":1,"group":3,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g3","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":2,"ep":2,"etp":1,"group":4,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":2,"ep":2,"etp":1,"group":5,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g5","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":2,"ep":2,"etp":1,"group":6,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g6","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":2,"ep":2,"etp":1,"group":7,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g7","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":0,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g0","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":1,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":2,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":3,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g3","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":4,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":5,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g5","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":6,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g6","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":2,"ep":1,"etp":1,"group":7,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g7","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":4,"ep":4,"etp":1,"group":0,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g0","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":4,"ep":4,"etp":1,"group":1,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":4,"ep":4,"etp":1,"group":2,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":4,"ep":4,"etp":1,"group":3,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g3","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":4,"ep":4,"etp":1,"group":4,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":4,"ep":4,"etp":1,"group":5,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g5","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":4,"ep":4,"etp":1,"group":6,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g6","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep-ellavox","cell":"cal-ellavox","cp":4,"ep":4,"etp":1,"group":7,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4|g7","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":0,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g0","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":1,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":2,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":3,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g3","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":4,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":5,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g5","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":6,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g6","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ellavox","cell":"cal-ellavox","cp":4,"ep":1,"etp":1,"group":7,"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g7","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-grpo-g16","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-grpo-g16","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-grpo-g16","cp":2,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-grpo-g16","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-grpo-g16","cp":4,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-grpo-g16","cp":4,"ep":4,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-grpo-g16","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":2}, + {"campaign":"tr-cost-l35-ep","cell":"cal-grpo-g16","cp":1,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":2}, + {"campaign":"tr-cost-l35-synth","cell":"cal-grpo-g4x4","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-grpo-g4x4","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-grpo-g4x4","cp":2,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-grpo-g4x4","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-grpo-g4x4","cp":4,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-grpo-g4x4","cp":4,"ep":4,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-grpo-g4x4","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":2}, + {"campaign":"tr-cost-l35-ep","cell":"cal-grpo-g4x4","cp":1,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":2}, + {"campaign":"tr-cost-l35-synth","cell":"cal-grpo-g8","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-grpo-g8","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-grpo-g8","cp":2,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-grpo-g8","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-grpo-g8","cp":4,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-grpo-g8","cp":4,"ep":4,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-grpo-g8","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":2}, + {"campaign":"tr-cost-l35-ep","cell":"cal-grpo-g8","cp":1,"ep":2,"etp":1,"group":null,"key":"cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":2}, + {"campaign":"tr-cost-l35-synth","cell":"cal-hetero2","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-hetero2","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-hetero2","cp":2,"ep":2,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-hetero2","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-hetero2","cp":4,"ep":2,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-hetero2","cp":4,"ep":4,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-hetero2","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":2}, + {"campaign":"tr-cost-l35-ep","cell":"cal-hetero2","cp":1,"ep":2,"etp":1,"group":null,"key":"cal-hetero2|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":2}, + {"campaign":"tr-cost-l35-synth","cell":"cal-hetero3","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-hetero3","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-hetero3","cp":2,"ep":2,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-hetero3","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-hetero3","cp":4,"ep":2,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-hetero3","cp":4,"ep":4,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-hetero3","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":2}, + {"campaign":"tr-cost-l35-ep","cell":"cal-hetero3","cp":1,"ep":2,"etp":1,"group":null,"key":"cal-hetero3|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":2}, + {"campaign":"tr-cost-l35-synth","cell":"cal-hetero","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-hetero","cp":2,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-hetero","cp":2,"ep":2,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-hetero","cp":4,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-hetero","cp":4,"ep":2,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-ep","cell":"cal-hetero","cp":4,"ep":4,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|ep4","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":1}, + {"campaign":"tr-cost-l35-synth","cell":"cal-hetero","cp":1,"ep":1,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":2}, + {"campaign":"tr-cost-l35-ep","cell":"cal-hetero","cp":1,"ep":2,"etp":1,"group":null,"key":"cal-hetero|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|ep2","layers":40,"model":"Qwen/Qwen3.5-35B-A3B","tp":2} + ], + "description": "Expected calibration cells for the Qwen3.5-35B-A3B (GDN + MoE, hidden 2,048) execution class on H200 bf16 (2026-09-04): the exact union of the shape-lattice launches below (dev/trainer_rank_cost_calibration_lattice.sky.yaml; CP separated from EP: EP1 at CP1/2/4 and TP2, EP2 at CP2/CP4/TP2, EP4 at CP4), keyed exactly as the fitter keys cells. tests/unit/test_planner_cost_manifest_recipes.py expands the launches and asserts they equal these cells. Exclusions are the cells this class cannot measure on one node, each with its reason.", + "excluded": [ + {"key":"cal-grpo-g16|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","reason":"cold memory admission lets a 106k-token forward run on one H200 and it hits CUDA out-of-memory in the expert grouped GEMM (issue #848)"}, + {"key":"cal-grpo-g4x4|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1","reason":"every candidate refused by memory admission on one H200 after the first observation (issue #848); no measurable rows"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g5","reason":"every candidate refused by memory admission on one H200 after the first observation (issue #848); no measurable rows"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g6","reason":"every candidate refused by memory admission on one H200 after the first observation (issue #848); no measurable rows"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp1|g7","reason":"94,721-token group: unsplit forward does not fit one H200 and forced layouts have no split-rung match; recorded as admission failures"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g2","reason":"EP1 with CP>1 on real-data routing segfaults in Transformer Engine's grouped GEMM (issue #851)"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|g7","reason":"every candidate refused by memory admission on one H200 after the first observation (issue #848); no measurable rows"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g0","reason":"EP1 with CP>1 on real-data routing segfaults in Transformer Engine's grouped GEMM (issue #851)"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g1","reason":"EP1 with CP>1 on real-data routing segfaults in Transformer Engine's grouped GEMM (issue #851)"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g2","reason":"EP1 with CP>1 on real-data routing segfaults in Transformer Engine's grouped GEMM (issue #851)"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g3","reason":"EP1 with CP>1 on real-data routing segfaults in Transformer Engine's grouped GEMM (issue #851)"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g4","reason":"EP1 with CP>1 on real-data routing segfaults in Transformer Engine's grouped GEMM (issue #851)"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g5","reason":"EP1 with CP>1 on real-data routing segfaults in Transformer Engine's grouped GEMM (issue #851)"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g6","reason":"EP1 with CP>1 on real-data routing segfaults in Transformer Engine's grouped GEMM (issue #851)"}, + {"key":"cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp4|g7","reason":"EP1 with CP>1 on real-data routing segfaults in Transformer Engine's grouped GEMM (issue #851)"} + ], + "launches": {"tr-cost-l35-ellavox":{"cells":["cal-ellavox:0:0","cal-ellavox:0:1","cal-ellavox:0:2","cal-ellavox:0:3","cal-ellavox:0:4","cal-ellavox:0:5","cal-ellavox:0:6","cal-ellavox:0:7"],"model":"Qwen/Qwen3.5-35B-A3B","recipe":"dev/trainer_rank_cost_calibration_lattice.sky.yaml","repeat":8,"shapes":[[1,1,1,1],[1,2,1,1],[1,4,1,1]]},"tr-cost-l35-ep":{"cells":["cal-grpo-g8:0","cal-grpo-g16:0","cal-grpo-g4x4:0","cal-hetero:0","cal-hetero2:0","cal-hetero3:0"],"model":"Qwen/Qwen3.5-35B-A3B","recipe":"dev/trainer_rank_cost_calibration_lattice.sky.yaml","repeat":8,"shapes":[[1,2,2,1],[1,4,2,1],[1,4,4,1],[2,1,2,1]]},"tr-cost-l35-ep-ellavox":{"cells":["cal-ellavox:0:0","cal-ellavox:0:1","cal-ellavox:0:2","cal-ellavox:0:3","cal-ellavox:0:4","cal-ellavox:0:5","cal-ellavox:0:6","cal-ellavox:0:7"],"model":"Qwen/Qwen3.5-35B-A3B","recipe":"dev/trainer_rank_cost_calibration_lattice.sky.yaml","repeat":8,"shapes":[[1,2,2,1],[1,4,4,1]]},"tr-cost-l35-synth":{"cells":["cal-grpo-g8:0","cal-grpo-g16:0","cal-grpo-g4x4:0","cal-hetero:0","cal-hetero2:0","cal-hetero3:0"],"model":"Qwen/Qwen3.5-35B-A3B","recipe":"dev/trainer_rank_cost_calibration_lattice.sky.yaml","repeat":8,"shapes":[[1,1,1,1],[1,2,1,1],[1,4,1,1],[2,1,1,1]]}}, + "recipes": [ + + ], + "schema": "art.dev.trainer_rank_cost_calibration_manifest.v1", + "table_id": "gdn-moe-h2048-h200-bf16" +} diff --git a/dev/trainer_rank_cost_fit.py b/dev/trainer_rank_cost_fit.py index d19bc436c..e2633e91d 100644 --- a/dev/trainer_rank_cost_fit.py +++ b/dev/trainer_rank_cost_fit.py @@ -87,6 +87,11 @@ class Candidate: # Grouping keys for per-model / per-shape reports (not fitted). model: str = "" shape: tuple[int, int, int, int] = (1, 1, 1, 1) + # Structure of the CP plan the layout produces (wave_count, + # max_rank_tokens, summary_ms) and the version-1 score: what the two-stage + # re-ranker prices and what it is compared against. Absent in older rows. + cp_plan: dict[str, float] | None = None + fallback_work: int | None = None def _shape(row: dict[str, Any]) -> tuple[int, int, int, int]: @@ -234,7 +239,11 @@ def cell_fingerprints(paths: list[Path]) -> dict[str, set[tuple[Any, ...]]]: def validate_manifest( - paths: list[Path], manifest_path: Path, *, excluded: list[str] + paths: list[Path], + manifest_path: Path, + *, + excluded: list[str], + exact: bool = False, ) -> tuple[list[str], list[str]]: """Exact cell identities: every expected cell present unless excluded, no unexpected cells, exclusions listed in the manifest, and one execution @@ -247,7 +256,11 @@ def validate_manifest( listed_exclusions = {cell["key"] for cell in manifest["excluded"]} present = cell_fingerprints(paths) problems: list[str] = [] - excluded_keys = {key for key in expected if any(p in key for p in excluded)} + excluded_keys = { + key + for key in expected + if (key in excluded if exact else any(p in key for p in excluded)) + } for key in sorted(excluded_keys - listed_exclusions): problems.append(f"excluded cell is not listed in the manifest: {key}") for key in sorted(expected - excluded_keys - set(present)): @@ -276,6 +289,8 @@ def export_certificate( report: dict[str, Any], manifest: dict[str, Any] | None = None, table_id: str = "", + reranked: list[Candidate] | None = None, + reranker: dict[str, Any] | None = None, ) -> None: """Write the compact, reproducible record binding the table to its data. @@ -308,7 +323,7 @@ def export_certificate( "shape": list(_shape(row)), } by_cell: dict[str, list[Candidate]] = defaultdict(list) - for candidate in candidates: + for candidate in [*candidates, *(reranked or [])]: by_cell[candidate.cell].append(candidate) cells = [ { @@ -322,6 +337,12 @@ def export_certificate( "median_ms": c.ms, "n": c.n, "spread_pct": c.spread_pct, + **({"cp_plan": c.cp_plan} if c.cp_plan is not None else {}), + **( + {"fallback_work": c.fallback_work} + if c.fallback_work is not None + else {} + ), } for c in sorted(members, key=lambda c: c.label) ], @@ -367,6 +388,9 @@ def unique(values: Any) -> list[Any]: "cells": cells, "integer_table_milli_us": integer_table, "integer_table_sha256": table_hash, + # Two-stage selection on the re-ranked shapes (None when the table + # scores every admitted shape directly). + "reranker": reranker, "metrics": { split: { k: report[split][k] @@ -429,11 +453,13 @@ def load_certificate(path: Path) -> tuple[list[Candidate], dict[str, Any]]: float(c["spread_pct"]), str(cell.get("model") or cell["cell"].split("|")[1]), tuple(cell.get("shape") or _shape(cell["facts"])), + c.get("cp_plan"), + c.get("fallback_work"), ) for cell in payload["cells"] for c in cell["candidates"] ] - return candidates, payload + return sorted(candidates, key=lambda c: (c.cell, c.label)), payload def load_candidates(paths: list[Path]) -> list[Candidate]: @@ -488,9 +514,14 @@ def load_candidates(paths: list[Path]) -> list[Candidate]: spread, str(cell["model"]), shape, + candidate.get("cp_plan"), + candidate.get("fallback_work"), ) ) - return candidates + # Canonical order: the regret refinement is a deterministic coordinate + # search whose path depends on candidate order, and equivalent optima + # exist, so evidence-file order and certificate order must agree. + return sorted(candidates, key=lambda c: (c.cell, c.label)) def evaluate_groups( @@ -590,8 +621,30 @@ def selector_check(candidates: list[Candidate]) -> dict[str, dict[str, Any]]: return report +def _cp_plan_value(candidate: Candidate, key: str) -> float: + if candidate.cp_plan is None: + raise ValueError( + f"{candidate.cell} {candidate.label}: no cp_plan recorded (re-run the " + "harness or backfill the evidence before fitting a re-ranker)" + ) + return float(candidate.cp_plan[key]) + + +# The re-ranker's terms, in the units ReRanker.score multiplies by layers: +# remote wave count and largest per-rank token load of the CP plan. +RERANK_TERMS: dict[str, Callable[[Candidate], float]] = { + "wave_per_layer": lambda c: ( + float(c.facts["layers"]) * _cp_plan_value(c, "wave_count") + ), + "max_rank_token_per_layer": lambda c: ( + float(c.facts["layers"]) * _cp_plan_value(c, "max_rank_tokens") + ), +} + + def term_matrix(candidates: list[Candidate], terms: tuple[str, ...]) -> np.ndarray: - """Term values in feature units (the integer functions divided by WORK_PER_US).""" + """Term values in feature units (the integer functions divided by + WORK_PER_US; the re-ranker's terms in their own units).""" rows = [] for candidate in candidates: @@ -607,7 +660,14 @@ def term_matrix(candidates: list[Candidate], terms: tuple[str, ...]) -> np.ndarr layers=int(candidate.facts["layers"]), gdn_layers=int(candidate.facts["gdn_layers"]), ) - rows.append([TERMS[name](features, facts) / WORK_PER_US for name in terms]) + rows.append( + [ + RERANK_TERMS[name](candidate) + if name in RERANK_TERMS + else TERMS[name](features, facts) / WORK_PER_US + for name in terms + ] + ) return np.asarray(rows, dtype=np.float64) @@ -768,6 +828,63 @@ def search(beta: np.ndarray, order: list[int]) -> tuple[float, np.ndarray]: return min(results, key=lambda item: item[0])[1] +def integerize( + candidates: list[Candidate], + terms: tuple[str, ...], + beta: np.ndarray, + *, + sweeps: int = 12, +) -> dict[str, int]: + """The production table: integer milli-microseconds per unit, repaired on + the integer grid when rounding costs a ranking. + + Rounding the float optimum can flip close pairs (equivalent optima with + tiny per-token coefficients are especially fragile). When the rounded + table's selection loss on the training cells is no worse than the float + optimum's, it is returned as is; otherwise a deterministic coordinate + search over integer values continues minimizing the same loss from the + rounded table, accepting only strict improvements. + """ + + matrix = term_matrix(candidates, terms) + table = np.asarray([int(round(value * 1_000)) for value in beta], dtype=np.int64) + + def loss(values: np.ndarray) -> float: + return selection_loss(candidates, matrix @ (values / 1_000.0)) + + best = loss(table) + if best <= selection_loss(candidates, matrix @ beta) + 1e-9: + return {name: int(value) for name, value in zip(terms, table, strict=True)} + for _ in range(sweeps): + improved = False + for j in range(len(table)): + current = int(table[j]) + trials = { + current + 1, + max(0, current - 1), + current + 2, + max(0, current - 2), + 0, + } + if current > 0: + trials |= { + int(round(current * factor)) for factor in (0.5, 0.9, 1.1, 2.0) + } + else: + trials |= {1, 10, 100, 1_000} + for trial in sorted(trials): + if trial == current or trial < 0: + continue + candidate_table = table.copy() + candidate_table[j] = trial + candidate_loss = loss(candidate_table) + if candidate_loss < best - 1e-9: + best, table, improved = candidate_loss, candidate_table, True + if not improved: + break + return {name: int(value) for name, value in zip(terms, table, strict=True)} + + def predict( candidates: list[Candidate], terms: tuple[str, ...], beta: np.ndarray ) -> np.ndarray: @@ -850,6 +967,282 @@ def gates_pass(report: dict[str, Any]) -> list[str]: return problems +RERANK_TERM_NAMES = ("wave_per_layer", "max_rank_token_per_layer") + + +def two_stage_shortlist( + members: list[Candidate], + shortlist_us: np.ndarray, + *, + shortlist_size: int, + incumbent: str, +) -> list[Candidate]: + """The cheapest ``shortlist_size`` layouts of a cell under the shortlist + score plus the incumbent anchor, in cheap order (production's rule).""" + + ranked = [ + member + for _score, _label, member in sorted( + zip(shortlist_us, [m.label for m in members], members, strict=True), + key=lambda item: (item[0], item[1]), + ) + ] + shortlist = ranked[: max(1, shortlist_size)] + shortlist += [m for m in ranked if m.label == incumbent and m not in shortlist] + return shortlist + + +def evaluate_two_stage( + candidates: list[Candidate], + *, + terms: tuple[str, ...], + shortlist_beta: np.ndarray, + rerank_beta: np.ndarray, + shortlist_size: int, + incumbent: str, +) -> dict[str, Any]: + """Two-stage selection per cell: shortlist by the ten-term score, select + the lowest re-ranker score (ties keep the cheaper layout). Reports the + ranking gates over the shortlist, shortlist recall and its irrecoverable + regret, non-regression against the cheap-only and version-1 selections, + and the synchronous planning cost of pricing the shortlist.""" + + by_cell: dict[str, list[Candidate]] = defaultdict(list) + for candidate in candidates: + by_cell[candidate.cell].append(candidate) + regrets: list[float] = [] + clear_misses: list[str] = [] + ordered_pairs = 0 + ordered_correct = 0 + recall = { + "cells": 0, + "best_in_shortlist": 0, + "clear_winner_cells": 0, + "clear_winners_in_shortlist": 0, + "max_irrecoverable_pct": 0.0, + } + regression: dict[str, Any] = { + "worse_than_cheap_cells": [], + "worse_than_fallback_cells": [], + "fallback_available": True, + "savings_vs_cheap_pct": [], + "savings_vs_fallback_pct": [], + } + economics: list[tuple[float, float, str]] = [] + per_cell: dict[str, dict[str, Any]] = {} + for cell, members in sorted(by_cell.items()): + shortlist_us = predict(members, terms, shortlist_beta) + shortlist = two_stage_shortlist( + members, shortlist_us, shortlist_size=shortlist_size, incumbent=incumbent + ) + rerank_us = dict( + zip( + [m.label for m in shortlist], + predict(shortlist, RERANK_TERM_NAMES, rerank_beta), + strict=True, + ) + ) + selected = min( + enumerate(shortlist), key=lambda item: (rerank_us[item[1].label], item[0]) + )[1] + best = min(members, key=lambda m: m.ms) + regret = (selected.ms - best.ms) / best.ms * 100.0 + regrets.append(regret) + by_ms = sorted(members, key=lambda m: m.ms) + lead_pct = (by_ms[1].ms - best.ms) / best.ms * 100.0 if len(by_ms) > 1 else 0.0 + clear = lead_pct > max(best.spread_pct, NOISE_BAND_PCT) + if clear and regret > 5.0: + clear_misses.append(cell) + recall["cells"] += 1 + recall["best_in_shortlist"] += best in shortlist + recall["clear_winner_cells"] += clear + recall["clear_winners_in_shortlist"] += clear and best in shortlist + irrecoverable = (min(m.ms for m in shortlist) - best.ms) / best.ms * 100.0 + recall["max_irrecoverable_pct"] = max( + recall["max_irrecoverable_pct"], irrecoverable + ) + for a, b in itertools.combinations(shortlist, 2): + separation = abs(a.ms - b.ms) / min(a.ms, b.ms) * 100.0 + if separation <= 3.0: + continue + ordered_pairs += 1 + ordered_correct += (a.ms < b.ms) == ( + rerank_us[a.label] < rerank_us[b.label] + ) + cheap = shortlist[0] + cheap_regret = (cheap.ms - best.ms) / best.ms * 100.0 + regression["savings_vs_cheap_pct"].append(cheap_regret - regret) + if regret - cheap_regret > 2.0: + regression["worse_than_cheap_cells"].append(cell) + if all(m.fallback_work is not None for m in members): + fallback = min(members, key=lambda m: (m.fallback_work or 0, m.label)) + fallback_regret = (fallback.ms - best.ms) / best.ms * 100.0 + regression["savings_vs_fallback_pct"].append(fallback_regret - regret) + if regret - fallback_regret > 2.0: + regression["worse_than_fallback_cells"].append(cell) + else: + regression["fallback_available"] = False + fallback_regret = None + summary_ms = [ + float((m.cp_plan or {}).get("summary_ms", 0.0)) for m in shortlist + ] + planning_ms = sum(summary_ms) - min( + summary_ms + ) # the selected plan is built anyway + economics.append((planning_ms / best.ms * 100.0, planning_ms, cell)) + per_cell[cell] = { + "selected": selected.label, + "best": best.label, + "regret_pct": regret, + "best_lead_pct": lead_pct, + "shortlist": [m.label for m in shortlist], + "cheap_regret_pct": cheap_regret, + "fallback_regret_pct": fallback_regret, + "irrecoverable_pct": irrecoverable, + "planning_ms": planning_ms, + } + regrets_sorted = sorted(regrets) + ratios = sorted(economics) + for key in ("savings_vs_cheap_pct", "savings_vs_fallback_pct"): + values = regression.pop(key) + regression[key.replace("_pct", "_mean_pct")] = ( + statistics.mean(values) if values else 0.0 + ) + return { + "cells": len(by_cell), + "pairwise_accuracy": (ordered_correct / ordered_pairs) + if ordered_pairs + else float("nan"), + "ordered_pairs": ordered_pairs, + "median_regret_pct": statistics.median(regrets) if regrets else float("nan"), + "p95_regret_pct": regrets_sorted[ + min(len(regrets_sorted) - 1, int(0.95 * len(regrets_sorted))) + ] + if regrets + else float("nan"), + "max_regret_pct": max(regrets) if regrets else float("nan"), + "clear_misses": clear_misses, + "recall": recall, + "non_regression": regression, + "planning_cost": { + "mean_pct_of_cell": statistics.mean(r[0] for r in ratios) + if ratios + else 0.0, + "median_pct_of_cell": ratios[len(ratios) // 2][0] if ratios else 0.0, + "max_pct_of_cell": ratios[-1][0] if ratios else 0.0, + "max_ms": ratios[-1][1] if ratios else 0.0, + "max_cell": ratios[-1][2] if ratios else "", + }, + "per_cell": per_cell, + } + + +def two_stage_gates(report: dict[str, Any]) -> list[str]: + problems = gates_pass(report) + recall = report["recall"] + if recall["clear_winners_in_shortlist"] != recall["clear_winner_cells"]: + problems.append( + f"shortlist misses clear winners: {recall['clear_winners_in_shortlist']}/{recall['clear_winner_cells']}" + ) + regression = report["non_regression"] + if regression["worse_than_cheap_cells"]: + problems.append( + f"worse than the cheap selection by >2%: {regression['worse_than_cheap_cells']}" + ) + if regression["worse_than_fallback_cells"]: + problems.append( + f"worse than the version-1 selection by >2%: {regression['worse_than_fallback_cells']}" + ) + # A synchronous miss pays for the shortlist's plans; on the re-ranked + # cells the execution saved over both single-stage selections must cover it. + cost = report["planning_cost"]["mean_pct_of_cell"] + savings = max( + regression["savings_vs_cheap_mean_pct"], + regression["savings_vs_fallback_mean_pct"], + ) + if cost > savings: + problems.append( + f"planning cost {cost:.1f}% of cell time exceeds the mean saving {savings:.1f}%" + ) + return problems + + +def fit_reranker( + candidates: list[Candidate], + reranked: list[Candidate], + *, + terms: tuple[str, ...], + held_out: Callable[[str], bool], + shortlist_size: int, + incumbent: str, + objective: str, +) -> dict[str, Any]: + """Fit and evaluate the two-stage selector for the re-ranked shapes. + + The shortlist score is the ten-term table fitted on every measured shape + (its job is recall, not selection); the re-ranker is fitted on paired + deltas within the training cells' shortlists and refined on their regret, + then integerized like a table. + """ + + shortlist_train = [c for c in candidates if not held_out(c.cell)] + shortlist_beta = nnls(*paired_deltas(shortlist_train, terms)) + if objective == "regret": + shortlist_beta = fit_regret(shortlist_train, terms, shortlist_beta) + shortlist_table = integerize(shortlist_train, terms, shortlist_beta) + shortlist_beta_int = np.asarray([shortlist_table[name] / 1_000.0 for name in terms]) + + by_cell: dict[str, list[Candidate]] = defaultdict(list) + for candidate in reranked: + by_cell[candidate.cell].append(candidate) + train_shortlists: list[Candidate] = [] + for cell, members in sorted(by_cell.items()): + if held_out(cell): + continue + train_shortlists += two_stage_shortlist( + members, + predict(members, terms, shortlist_beta_int), + shortlist_size=shortlist_size, + incumbent=incumbent, + ) + rerank_beta = nnls(*paired_deltas(train_shortlists, RERANK_TERM_NAMES)) + if objective == "regret": + rerank_beta = fit_regret(train_shortlists, RERANK_TERM_NAMES, rerank_beta) + rerank_table = integerize(train_shortlists, RERANK_TERM_NAMES, rerank_beta) + rerank_beta_int = np.asarray( + [rerank_table[name] / 1_000.0 for name in RERANK_TERM_NAMES] + ) + + def two_stage(subset: list[Candidate]) -> dict[str, Any]: + return evaluate_two_stage( + subset, + terms=terms, + shortlist_beta=shortlist_beta_int, + rerank_beta=rerank_beta_int, + shortlist_size=shortlist_size, + incumbent=incumbent, + ) + + held = [c for c in reranked if held_out(c.cell)] + shapes = sorted({_shape_label(c) for c in reranked}) + return { + "shortlist_size": shortlist_size, + "incumbent": incumbent, + "shapes": shapes, + "shortlist_table_milli_us": shortlist_table, + "wave_per_layer_milli_us": rerank_table["wave_per_layer"], + "max_rank_token_per_layer_milli_us": rerank_table["max_rank_token_per_layer"], + "metrics": { + "all": two_stage(reranked), + "held_out": two_stage(held) if held else None, + "by_shape": { + shape: two_stage([c for c in reranked if _shape_label(c) == shape]) + for shape in shapes + }, + }, + } + + def main() -> None: parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("evidence", nargs="*", type=Path) @@ -868,7 +1261,8 @@ def main() -> None: default="", help=( "comma-separated substrings of cells to leave out explicitly (recorded in the " - "certificate); the fitter never drops incomplete cells silently" + "certificate), or @manifest for exactly the manifest's listed exclusions; " + "the fitter never drops incomplete cells silently" ), ) parser.add_argument( @@ -898,6 +1292,17 @@ def main() -> None: help="identity of the production table this certificate binds (required with --export-certificate)", ) parser.add_argument("--integerize", action="store_true") + parser.add_argument( + "--rerank-shapes", + default="", + help=( + "comma-separated shape labels (e.g. tp1cp4) selected by the two-stage " + "re-ranker: the ten-term score shortlists, the CP plan structure decides; " + "their cells leave the direct fit and need cp_plan in the evidence" + ), + ) + parser.add_argument("--shortlist-size", type=int, default=3) + parser.add_argument("--incumbent", default="depth_one") parser.add_argument( "--objective", choices=("lsq", "regret"), @@ -911,15 +1316,27 @@ def main() -> None: ) arguments = parser.parse_args() terms = tuple(t for t in arguments.terms.split(",") if t) - excluded = [p for p in arguments.exclude_cells.split(",") if p] + if arguments.exclude_cells == "@manifest": + # Exactly the manifest's listed exclusions (each carries its reason). + if not arguments.manifest: + raise SystemExit("--exclude-cells @manifest requires --manifest") + manifest_excluded = json.loads(Path(arguments.manifest).read_text())["excluded"] + excluded = [cell["key"] for cell in manifest_excluded] + exact = True + else: + excluded = [p for p in arguments.exclude_cells.split(",") if p] + exact = False def excluded_cell(cell: str) -> bool: - return any(p in cell for p in excluded) + return cell in excluded if exact else any(p in cell for p in excluded) manifest_record: dict[str, Any] | None = None if arguments.manifest: problems, excluded_keys = validate_manifest( - arguments.evidence, Path(arguments.manifest), excluded=excluded + arguments.evidence, + Path(arguments.manifest), + excluded=excluded, + exact=exact, ) if problems: print("manifest validation failed:", file=sys.stderr) @@ -972,6 +1389,27 @@ def held_out(cell: str) -> bool: return int(cell.rsplit("|g", 1)[1]) % 2 == 1 return False + # Cells on re-ranked shapes are selected by the two-stage selector; the + # direct table is fitted and gated on the other shapes only. + rerank_shapes = {label for label in arguments.rerank_shapes.split(",") if label} + reranked = [c for c in candidates if _shape_label(c) in rerank_shapes] + all_candidates = candidates + if rerank_shapes: + if not reranked: + raise SystemExit( + f"no cells on the re-ranked shapes {sorted(rerank_shapes)}" + ) + candidates = [c for c in candidates if _shape_label(c) not in rerank_shapes] + if not candidates: + raise SystemExit( + "every cell is on a re-ranked shape: the direct table needs directly " + "scored cells (for a validation run of the two-stage selection alone, " + "read the timed `automatic` rows without --rerank-shapes)" + ) + print( + f"re-ranked shapes {sorted(rerank_shapes)}: {len({c.cell for c in reranked})} cells " + f"leave the direct fit ({len({c.cell for c in candidates})} direct cells)" + ) train = [c for c in candidates if not held_out(c.cell)] test = [c for c in candidates if held_out(c.cell)] beta = nnls(*paired_deltas(train, terms)) @@ -986,10 +1424,10 @@ def held_out(cell: str) -> bool: "fit_all": evaluate(candidates, predict(candidates, terms, beta)), } if arguments.integerize: - # Production stores integer milli-microseconds per feature unit. - integer = { - name: int(round(value * 1_000)) for name, value in report["terms"].items() - } + # Production stores integer milli-microseconds per feature unit; + # repaired on the integer grid (training cells) when rounding costs + # a ranking. + integer = integerize(train, terms, beta) report["integer_terms_milli_us"] = integer beta_int = np.asarray( [integer[name] / 1_000.0 for name in terms], dtype=np.float64 @@ -998,8 +1436,66 @@ def held_out(cell: str) -> bool: candidates, predict(candidates, terms, beta_int) ) print("COEFFICIENTS_MILLI_US =", json.dumps(integer, indent=4)) + if reranked: + if not arguments.integerize: + raise SystemExit("--rerank-shapes requires --integerize") + report["reranker"] = fit_reranker( + all_candidates, + reranked, + terms=terms, + held_out=held_out, + shortlist_size=arguments.shortlist_size, + incumbent=arguments.incumbent, + objective=arguments.objective, + ) + block = report["reranker"] + print( + "RERANKER =", + json.dumps( + { + k: block[k] + for k in ( + "shortlist_size", + "incumbent", + "shapes", + "wave_per_layer_milli_us", + "max_rank_token_per_layer_milli_us", + ) + }, + indent=4, + ), + ) + print( + "shortlist table (milli-us):", json.dumps(block["shortlist_table_milli_us"]) + ) + for name, metrics in ( + ("two-stage all", block["metrics"]["all"]), + ("two-stage held-out", block["metrics"]["held_out"]), + *( + (f"two-stage {shape}", m) + for shape, m in block["metrics"]["by_shape"].items() + ), + ): + if not metrics: + continue + recall = metrics["recall"] + problems = two_stage_gates(metrics) + print( + f"{name:22s} cells={metrics['cells']:3d} acc={metrics['pairwise_accuracy']:.3f} regret median={metrics['median_regret_pct']:.2f}% " + f"p95={metrics['p95_regret_pct']:.2f}% max={metrics['max_regret_pct']:.2f}% | recall {recall['best_in_shortlist']}/{recall['cells']} " + f"clear {recall['clear_winners_in_shortlist']}/{recall['clear_winner_cells']} irrecoverable max {recall['max_irrecoverable_pct']:.1f}% | " + f"planning mean {metrics['planning_cost']['mean_pct_of_cell']:.1f}% max {metrics['planning_cost']['max_pct_of_cell']:.1f}% of cell vs saving " + f"{max(metrics['non_regression']['savings_vs_cheap_mean_pct'], metrics['non_regression']['savings_vs_fallback_mean_pct']):.1f}% | " + + ( + "GATES PASS" + if not problems + else "GATES FAIL: " + "; ".join(problems) + ) + ) report["production"] = ( - production_regret(candidates, arguments.evidence) if arguments.evidence else {} + production_regret(all_candidates, arguments.evidence) + if arguments.evidence + else {} ) if report["production"]: regrets = sorted(v["regret_pct"] for v in report["production"].values()) @@ -1083,11 +1579,16 @@ def held_out(cell: str) -> bool: "holdout": arguments.holdout, "exclude_cells": excluded, "require_complete": arguments.require_complete, + "rerank_shapes": sorted(rerank_shapes), + "shortlist_size": arguments.shortlist_size, + "incumbent": arguments.incumbent, }, integer_table=report["integer_terms_milli_us"], report=report, manifest=manifest_record, table_id=arguments.table_id, + reranked=reranked, + reranker=report.get("reranker"), ) print("certificate written:", arguments.export_certificate) diff --git a/dev/trainer_rank_landing_acceptance.py b/dev/trainer_rank_landing_acceptance.py index af2a9b044..c0de5ea97 100644 --- a/dev/trainer_rank_landing_acceptance.py +++ b/dev/trainer_rank_landing_acceptance.py @@ -59,6 +59,7 @@ from pathlib import Path import statistics import sys +import time from typing import Any, cast REPO_ROOT = Path(__file__).resolve().parents[1] @@ -1131,6 +1132,31 @@ def _gather_objects(value: Any, group: Any = None) -> list[Any]: return values +def _hybridep_telemetry(rank: Any) -> dict[str, Any]: + """HybridEP facts of the last forward on MoE runtimes under expert + parallelism: the largest per-rank row count TrainerRank configured (its + high-water mark) and the live buffer's token capacity per rank. Both are + ``None`` without expert parallelism.""" + + try: + from megatron.core import parallel_state as ps + from megatron.core.transformer.moe import fused_a2a + except ImportError: + return {"hybridep_rows_high_water": None, "hybridep_capacity_rows": None} + if int(ps.get_expert_model_parallel_world_size()) <= 1: + return {"hybridep_rows_high_water": None, "hybridep_capacity_rows": None} + buffer = getattr(fused_a2a, "_hybrid_ep_buffer", None) + capacity = ( + int(buffer.configurer.buffer_config.max_num_of_tokens_per_rank) + if buffer is not None + else None + ) + return { + "hybridep_rows_high_water": int(getattr(rank, "_hybridep_rows_high_water", 0)), + "hybridep_capacity_rows": capacity, + } + + def _merge_rank_statuses(gathered: list[list[str]]) -> list[str]: return sorted({str(status) for statuses in gathered for status in statuses}) @@ -2097,10 +2123,12 @@ def phase_cost_calibrate( from art.megatron import train as megatron_train from art.trainer_rank import TrainerRank, TrainerRankMemoryError from art.trainer_rank._planner_cost import ( + COEFFICIENT_VERSION_FALLBACK, DeviceClass, ScoringFacts, layout_features, predicted_us, + prefix_tree_layout_score, ) from art.trainer_rank._prefix_tree_planner import ( build_canonical_prefix_tree, @@ -2197,15 +2225,45 @@ def current_score(features: Any) -> float | None: return None return predicted_us(features, scoring_facts, rank._planner_coefficients) + rows_for_plan = tuple( + r.input_tokens.reshape(-1).to(torch.long) for r in requests + ) for candidate in candidates: features = layout_features(candidate.layout) current_us = current_score(features) + # The CP plan structure the layout produces (what the two-stage + # re-ranker prices) and what deriving it costs on this host. + started = time.perf_counter() + if facts.cp_size > 1: + wave_count, max_rank_tokens = rank._plan_structure( + rows_for_plan, tree, candidate.layout + ) + else: + wave_count, max_rank_tokens = 0, int(candidate.layout.packed_tokens) candidate_rows.append( { "label": candidate.labels[0], "labels": list(candidate.labels), "features": features.as_dict(), "current_score_us": current_us, + "cp_plan": { + "wave_count": int(wave_count), + "max_rank_tokens": int(max_rank_tokens), + "summary_ms": (time.perf_counter() - started) * 1_000.0, + }, + # The version-1 score: the fallback the two-stage + # certification must not regress against. + "fallback_work": int( + prefix_tree_layout_score( + candidate.layout, + cp_size=facts.cp_size, + layers=facts.layers, + uses_gdn=facts.uses_gdn, + tp_size=facts.tp_size, + gdn_layers=facts.gdn_layers, + coefficient_version=COEFFICIENT_VERSION_FALLBACK, + )[0] + ), } ) # The production selector's own choice, timed like every other @@ -2274,6 +2332,16 @@ def run(label: str) -> dict[str, object]: except TrainerRankMemoryError as error: failed = 1 message = str(error) + except ValueError as error: + # The unsplit forward did not fit and admission tried a split + # rung; the forced candidate label names a layout of the whole + # tree, so it has no match in a sub-forward's family. A split + # run is not a comparable measurement anyway (the fitter drops + # subforward_count > 1), so record it as an admission failure. + if "unknown forced layout anchor" not in str(error): + raise + failed = 1 + message = f"forced layout has no match under a split rung: {error}" end.record() torch.cuda.synchronize() flags = torch.tensor([float(failed)], device="cuda") @@ -2296,6 +2364,7 @@ def run(label: str) -> dict[str, object]: "subforward_count": int(telemetry["subforward_count"]), "peak_allocated_bytes": int(torch.cuda.max_memory_allocated()), "loss": float(loss.detach().float().item()), + **_hybridep_telemetry(rank), } del outputs, loss rank.zero_grad() diff --git a/dev/trainer_rank_landing_acceptance_README.md b/dev/trainer_rank_landing_acceptance_README.md index ea28a83b1..890b16867 100644 --- a/dev/trainer_rank_landing_acceptance_README.md +++ b/dev/trainer_rank_landing_acceptance_README.md @@ -22,8 +22,14 @@ every gate below now passes on the landed implementation. | `--phase dp2-tp2-waves` (`dev/trainer_rank_landing_acceptance_dp2_tp2.sky.yaml`) | 4x H200 k8s | DP2×TP2 public `forward_micro_batches`: ≥2 waves, distinct DP payloads, identical wave shapes within each TP pair, every input returned once in order, forward+backward per wave, automatic vs depth-one parity, empty-DP-slot arm, no hang | see evidence log | | `--phase cost-calibrate` (`dev/trainer_rank_cost_calibration_{cp4,2gpu}.sky.yaml`, `dev/trainer_rank_cost_calibration_local.sh`) | 1x/2x/4x H200 | every mandatory candidate layout of a cell timed through the public API (forward+backward, active LoRA, compile-free, max-rank) plus the production selection; layout features and topology/model facts to JSONL. Cells: GRPO g8/g16/g4x4 (Qwen3.5-4B GDN and Qwen3-4B attention, 2-layer and full height), three heterogeneous controls, Ellavox groups, at TP1/TP2 × CP1/CP2/CP4 | 58 cells, 3,849 within-cell pairs | | `dev/trainer_rank_cost_fit.py` | CPU | paired within-cell deltas, non-negative least squares over the production term functions, regret-minimizing refinement; gates on whole held-out cells: pairwise ordering ≥90% on pairs separated >3%, median regret ≤2%, p95 ≤5%, none >10%, clear winners selected within 5%; `--selector-check` runs the shipped table through the real selector | final table (fit on 45 cells, evaluated on all 58; 3,849 pairs; 13 held-out cells): 98.1% pairwise, p95 regret 2.9%, max 4.2%, no clear misses; pre-registered odd-Ellavox holdout passes; the two Ellavox CP4 cells re-measured after the issue #840 harness fix are held out too (shipped-table regret 0% and 2.8%); profile narrowed to the measured envelope (H200-class SM 9.0, bf16, hidden 2,560, dense) and bound to the certificate by test; expected-cell manifest validated; TP2, CP2 and attention-model ablations pass (all-heterogeneous ablation: one CP4 cell at 9.5%); campaign-1 table on the 18 later cells within 4.2%; timed production selection on the later CP2/TP2 cells: median −0.2%, max 0.4%; hand-set score: 78.6% pairwise, max regret 67% | -| `dev/trainer_rank_cost_calibration_certificate.json` (compact: one line per cell) + `tests/unit/test_planner_cost_certificate.py` | CPU | the shipped table is the certified table (hash) and the certified metrics hold on the recorded per-cell aggregates; exact 58 cell identities match the manifest with no exclusions; the admitted execution classes (device classes, dtypes, model geometries, parallel shapes) equal the certified cells' and every geometry was measured at every admitted shape; runners fail loudly on any cell failure and the fitter refuses incomplete evidence unless exclusions are explicit (the two Ellavox CP4 cells that hung in the campaign were a harness desynchronization, issue #840, fixed and re-measured) | pass | -| `tests/unit/test_planner_cost_profile.py` | CPU | a fitted table applies only to the execution classes it was measured on (device class, dtype, Megatron-config model geometry, TP × CP × EP × ETP shape), bound to its certificate by test; outside every table the version-1 score is used verbatim, unmeasured shapes such as TP2 × CP2 included | pass | +| `dev/trainer_rank_cost_calibration_lattice.sky.yaml` (`MODEL`, `SHAPES` = tp,cp,ep[,etp], `CELLS`) | 4x H200 | one model's calibration cells over an explicit parallel-shape lattice, one torchrun per shape with EP/ETP pinned (builds HybridEP in setup); used for the Qwen3.5-35B-A3B class (EP1 at CP1/2/4/TP2, EP2 at CP2/CP4/TP2, EP4 at CP4) and the dense controls | 35B class: 73 cells / 3,658 pairs certified as `gdn-moe-h2048-h200-bf16` (97.1% pairwise, p95 regret 2.4%, max 4.9%, gates pass per shape); 15 lattice cells excluded with reasons (#848, #851) | +| dense attention classes (`dev/trainer_rank_cost_calibration_lattice.sky.yaml`: Qwen3-1.7B, Qwen3-8B, Qwen3-14B; Qwen3-4B real data at CP4/TP2×CP2) | 4x H200 | same protocol on the Qwen3-tokenized Ellavox groups and synthetic families at TP1 × CP1/2/4, TP2 × CP1, TP2 × CP2 (70 cells each); candidate rows record the CP plan structure (`cp_plan`: wave count, max rank tokens, summary cost) and the version-1 score | `dense-attn-h2048`/`h4096` at CP1, CP2, TP2 (42 cells; pairwise 96.5% / 99.9%, max regret 1.0% / 1.1%), `dense-attn-h5120` also at TP2 × CP2 (56 cells; 99.9%, max 1.0%); TP1 × CP4 admitted through the two-stage re-ranker (`--rerank-shapes tp1cp4`: shortlist of 3 + depth-one anchor, re-rank by waves and max rank load; two-stage max regret 2.6% / 0.8% / 4.1%, every clear winner shortlisted, no >2% regression vs cheap or version 1, planning cost mean 2.6% / 1.6% / 1.0% of cell time below the mean saving); TP2 × CP2 not admitted for the two smaller classes; the dense table withholds Qwen3-4B × TP1 × CP4 (35% vs version 1's 15%) | +| Qwen3-30B-A3B class (`dev/trainer_rank_cost_calibration_lattice.sky.yaml`, EP1 at CP1/2/4/TP2, EP2 at CP2/CP4/TP2, EP4 at CP4; CP2 and CP4 EP1/EP4 re-measured with sixteen rounds) | 4x H200 | 112 cells on the Qwen3-tokenized Ellavox groups and synthetic families; 3 CP1 cells fail single-GPU memory admission (#848); sporadic multi-second stalls on ~4% of rounds, medians hold at sixteen rounds | `attn-moe-h2048-h200-bf16`: direct at CP1, CP2, CP2 × EP2, TP2, TP2 × EP2 (67 cells; 96.3%, max 2.2%); CP4 at EP1/EP2/EP4 through the re-ranker (42 cells; max 2.1%, planning 1.1% vs 1.4% saved); version 1 lost up to 13% at CP4 | +| Qwen3.5-27B class (`dev/trainer_rank_cost_calibration_lattice.sky.yaml`, TP1 × CP1/2/4, TP2 × CP1, TP2 × CP2 in two halves) | 4x H200 | 70 cells on the Qwen3.5-tokenized Ellavox groups and synthetic families; 4 cells incomplete under single-GPU memory admission (#848), 1 lost to an NCCL timeout | `dense-gdn-h5120-h200-bf16`: admitted at CP1, CP2, CP4, TP2 (52 cells; pairwise 97.0%, max regret 3.4%, held-out 1.8%) without a re-ranker; TP2 × CP2 fails its gates (p95 5.1%) and keeps version 1, which lost up to 112% at CP4 on this class | +| `dev/trainer_rank_cost_calibration_certificate_.json` (one per calibrated table, compact: one line per cell) + `tests/unit/test_planner_cost_certificate.py` | CPU | the shipped table is the certified table (hash) and the certified metrics hold on the recorded per-cell aggregates; exact cell identities match each table's manifest (dense: the 58 fitted cells plus the 16 excluded blind-spot cells; attention classes: their CP4 and failing TP2 × CP2 cells excluded with reasons) and each table's withheld pairs are exactly the manifest's; tables with a re-ranker: the shipped re-ranker equals the certificate's `reranker` block and the two-stage selection recomputed from the recorded aggregates reproduces its metrics and passes the gates on every re-ranked shape; the admitted execution classes (device classes, dtypes, model geometries, parallel shapes) equal the certified cells' and every geometry was measured at every admitted shape; runners fail loudly on any cell failure and the fitter refuses incomplete evidence unless exclusions are explicit (the two Ellavox CP4 cells that hung in the campaign were a harness desynchronization, issue #840, fixed and re-measured) | pass | +| `tests/unit/test_planner_rerank.py` | CPU | two-stage selection: shortlist (k + depth-one anchor), lowest second-stage score wins, ties keep the cheaper layout, tables validate re-ranked shapes against their re-ranker, `select_scoring` returns the re-ranker on re-ranked shapes only | pass | +| `tests/unit/test_context_parallel_plan_evaluation.py` | CPU (megatron env) | the vectorized CP assignment-search evaluator equals the scalar reference on randomized instances (incl. the long-row path) and the search result is unchanged | pass | +| `tests/unit/test_planner_cost_profile.py` | CPU | a fitted table applies only to the execution classes it was measured on (device class, dtype, Megatron-config model geometry, TP × CP × EP × ETP shape), bound to its certificate by test; outside every table the version-1 score is used verbatim, unmeasured shapes such as CP8 and withheld pairs (Qwen3-4B at TP1 × CP4) included | pass | | `--phase split-conversion --pressure ballast` | 1x H200 | same cell under real pressure (live ballast tensors, no test hooks). Training forward: ballast sized from the measured retained fraction f (budget midway inside the (1−f)·R/2 conversion window, width reported), unsplit refused before execution, split runs under the reduced headroom with parity, combined backward with ballast live, observed forward+backward peak within both the admitted budget and the predicted peak; deeper ballast refuses before execution. `no_grad` forward (retained ≈ outputs only; the demonstrated high-value case): budget 60% of the unsplit requirement converts with parity and the observed peak within budget and prediction | see evidence log | Protocol note (2026-09-01): planning cost is gated as a *fraction* only on the diff --git a/dev/trainer_rank_planner_design.md b/dev/trainer_rank_planner_design.md index eff392d2d..539f2459e 100644 --- a/dev/trainer_rank_planner_design.md +++ b/dev/trainer_rank_planner_design.md @@ -156,17 +156,117 @@ geometry read from the Megatron config (hidden and FFN widths, attention head geometry, GDN state shape, expert geometry — never a model name; layer counts are scoring facts, not identity) and parallel shape (TP × CP × EP × ETP). Admission is exact: the dense hidden-2,560 table admits the Qwen3.5-4B and -Qwen3-4B geometries at the four measured shapes (TP1 × CP1/2/4 and TP2 × CP1); -TP2 × CP2, CP8, TP4, expert parallelism, other widths and MoE models keep the -version-1 score (kept verbatim) with a one-time warning. Each certificate names -its table and records the admitted device classes, dtypes, geometries and -shapes; the certificate test asserts they equal the production table's sets -and that every geometry was measured at every admitted shape. -`dev/trainer_rank_cost_calibration_manifest.json` lists the exact cells each -recipe launches; the fitter's `--manifest` validation requires every +Qwen3-4B geometries at the four measured shapes (TP1 × CP1/2/4 and TP2 × CP1) +except the withheld pair Qwen3-4B × TP1 × CP4 (measured and fitted, but the +score is known to misrank real-data groups there; see the known limitation +below). A table's `withheld` pairs fall back to version 1, and the manifest +records each with its reason. TP2 × CP2, CP8, TP4, expert parallelism, other +widths and MoE models keep the version-1 score (kept verbatim) with a one-time +warning unless another table admits them. Each certificate names its table and +records the admitted device classes, dtypes, geometries and shapes; the +certificate test asserts they equal the production table's sets, that every +geometry was measured at every admitted shape, and that the withheld pairs are +exactly the manifest's. +`dev/trainer_rank_cost_calibration_manifest_
.json` lists the exact cells +each recipe or lattice launch of that table produces; the fitter's `--manifest` validation requires every non-excluded cell to be present and complete, rejects unexpected cells and duplicate cells with differing execution fingerprints (geometry included), and -the certificate test asserts all 58 identities (no exclusions). +the certificate test asserts the 58 fitted identities plus the 16 excluded +blind-spot cells of the Qwen3-4B real-data launch (below). + +Second calibrated class (2026-09-04): Qwen3.5-35B-A3B (GDN + MoE, hidden +2,048, 256 experts top-8) on H200 bf16, measured over a shape lattice that +separates context from expert parallelism (EP1 at CP1/2/4 and TP2, EP2 at +CP2/CP4/TP2, EP4 at CP4; `dev/trainer_rank_cost_calibration_lattice.sky.yaml`) +on the synthetic families and the Ellavox groups: 73 cells, 3,658 within-cell +pairs, 13 held-out odd-Ellavox cells. The same ten terms fitted to their own +table (`gdn-moe-h2048-h200-bf16`) rank 97.1% of separated pairs correctly, +median regret 0%, p95 2.4%, max 4.9%, gates passing on every shape (EP shapes +included), so no expert-parallel term was needed; the version-1 fallback this +class used before loses up to 35% on the large Ellavox groups. Fifteen cells +of the lattice are excluded with reasons: CP1 cells the single-GPU memory +admission cannot run or refuses (issue #848) and the CP4/EP1 Ellavox cells +that segfault in the expert grouped GEMM on real routing (issue #851). + +Known limitation (2026-09-04, dense controls): the O(segments) layout +features cannot see the context-parallel plan a layout produces, and on +attention-only models with real data at CP4 that plan dominates. Two layouts +of an Ellavox group with near-identical features (about 13k packed tokens, +six versus seven segments, the same 12k-token longest segment) differ by 25% +in measured time because one needs four exchange waves with rank loads +7168/3584/1536/790 and the other two waves with 6656/2560/2048/1751. The +Qwen3-1.7B, Qwen3-8B and Qwen3-14B controls therefore fail the gates at CP4 +with the ten terms (held-out max regret 26%, 14% and 17%), and the same group +measured on the certified Qwen3-4B geometry at TP1 × CP4 shows a 35% regret +for the shipped table (the version-1 score reaches 15% on another group +there; neither is adequate). The dense certificate had no +attention-plus-real-data cells at CP4, so its metrics did not cover this. + +Resolution (2026-09-04, per review). The attention classes get their own +tables where the ten-term gates pass — Qwen3-1.7B (`dense-attn-h2048-h200-bf16`) +and Qwen3-8B (`dense-attn-h4096-h200-bf16`) at TP1 × CP1/CP2 and TP2 × CP1, +Qwen3-14B (`dense-attn-h5120-h200-bf16`) also at TP2 × CP2 (42/42/56 cells; +pairwise 96.5%/99.9%/99.9%, max regret ≤1.1%) — and TP1 × CP4 is admitted +through a **two-stage selection** (`ReRanker`, `CalibratedTable.reranked_shapes`, +`select_prefix_tree_layout(..., reranker, plan_structure)`): the ten-term score +(a shortlist table fitted on every measured shape; its job is recall) keeps the +three cheapest layouts of the search plus the depth-one anchor, each shortlisted +layout is priced by the structure of the context-parallel plan it produces — +remote wave count and largest per-rank token load, per layer, from the CP +planner's own assignment (`summarize_prefix_tree_plan`, through the planning +bundle cache so the selected layout's plan is reused) — and the lowest +second-stage score wins (ties keep the cheaper layout). The two physical +drivers came out of the measured schedules: on equal-load layouts an extra +remote wave costs 2.4 / 1.8 / 0.6 ms per layer on 1.7B / 8B / 14B (the CP +planner's own model prefers the extra wave; issue #854), and the max-rank load +carries the rest. Certified per class on TP1 × CP4 (`reranker` block of the +certificate, bound by `tests/unit/test_planner_cost_certificate.py`): +every clear measured winner is in the shortlist; two-stage regret max +2.6% / 0.8% / 4.1% (held-out 0.7% / 0.8% / 0.4%); never more than 2% worse +than the cheap-only or the version-1 selection; and the synchronous planning +cost of pricing the shortlist (mean 2.6% / 1.6% / 1.0% of the cell's time +after the assignment-search speedup below) is below the mean execution saved +on those cells (2.9% / 1.8% / 1.2%). TP2 × CP2 still fails its gates for the +two smaller classes and keeps version 1; the dense table withholds +Qwen3-4B × TP1 × CP4 (version 1's 15% worst observed cell is the less harmful +fallback than the table's 35%; its 7 real-data CP4 cells are too few to +certify a re-ranker). + +Fourth calibrated class (2026-09-04/05): Qwen3-30B-A3B (attention + MoE, hidden +2,048, 128 experts top-8; `attn-moe-h2048-h200-bf16`) on H200 bf16 over the same +EP-deconfounding lattice (112 cells; 3 CP1 cells excluded for the single-GPU +memory admission of issue #848; no CP>1/EP1 segfaults on this model). The +eight-round timings of TP1 × CP2 (EP1/EP2) and TP1 × CP4 (EP1/EP4) carried +sporadic 10–20 s stalls on about 4% of rounds, so those shapes were re-measured +with sixteen rounds (the stalls persist at about 4% of rounds but shrink to +about 2×, and medians hold). The class shows the attention CP4 blind spot. +Admitted directly at TP1 × CP1, TP1 × CP2 (EP1, EP2) and TP2 × CP1 (EP1, EP2) +(67 cells; pairwise 96.3%, max regret 2.2%), and at TP1 × CP4 with EP1, EP2 and +EP4 through the two-stage re-ranker (42 cells; max 2.1%, every clear winner +shortlisted, planning 1.1% of cell time against 1.4% saved), where version 1 +lost up to 13%. + +Fifth calibrated class (2026-09-04): Qwen3.5-27B (dense GDN + attention, +hidden 5,120; `dense-gdn-h5120-h200-bf16`) on H200 bf16 over TP1 × CP1/2/4, +TP2 × CP1 and TP2 × CP2 (70 cells; four CP1/CP2 cells incomplete or lost to the +single-GPU memory admission of issue #848, one TP2 × CP2 cell lost to an NCCL +collective timeout during warm-up). Clean timings (median per-candidate spread +0.4–0.7% except TP2 × CP2), and the ten terms rank this GDN class at CP4 as they +do the 35B GDN MoE class, so no re-ranker: admitted directly at TP1 × CP1/2/4 +and TP2 × CP1 (52 cells; pairwise 97.0%, max regret 3.4%, held-out 1.8%). +TP2 × CP2 fails its gates marginally (p95 5.1%, one clear miss) and keeps +version 1. This is the class where the version-1 fallback hurt most: on the +largest Ellavox group at CP4 it chose a layout 112% slower than the best +(4,084 ms against 1,927 ms), and 26% slower at TP2 × CP2. + +The re-ranker is affordable because the context-parallel assignment search +was vectorized (`_evaluate_plans` in `art.megatron.context_parallel.runtime`): +it prices a whole batch of candidate moves in one numpy pass with the cost +formulas and stage simulations kept in the same arithmetic order, identical +results on all 446 real calibration candidates (the pre-rewrite evaluator is +the reference in `tests/unit/test_context_parallel_plan_evaluation.py`), and +the median search fell from 18.3 ms to 4.1 ms per packed row — also on the +production plan-build path of every CP>1 micro-batch. Landing gates re-derived: the sealed win-cell shape still selects deep sharing (prompt-level sharing at CP4, where it measures fastest; full sharing at CP1), @@ -184,7 +284,8 @@ measured widths. The selected table's identity joins the coefficient version in the planner facts and therefore in every layout cache key. CPU-only planning (unit tests) uses the default dense table. -Reproducibility (review finding): `dev/trainer_rank_cost_calibration_certificate.json` +Reproducibility (review finding): `dev/trainer_rank_cost_calibration_certificate_
.json` +(one per calibrated table) binds the shipped table to its data — per-cell candidate features, median timings, counts, spreads and fingerprints (no tokens, no per-sample rows), the exact fit arguments including any explicit cell exclusions, the integer table diff --git a/src/art/megatron/context_parallel/runtime.py b/src/art/megatron/context_parallel/runtime.py index b373e7c99..6eeaa3631 100644 --- a/src/art/megatron/context_parallel/runtime.py +++ b/src/art/megatron/context_parallel/runtime.py @@ -7,6 +7,7 @@ from threading import Lock from typing import Any, cast +import numpy as np from pydantic import BaseModel import torch @@ -37,7 +38,6 @@ TrainingMicrobatchWorkload, ) -_CHUNK_MASK_STATS_TORCH_THRESHOLD = 1024 _CP4_SEARCH_PROBE_CANDIDATE_LIMIT = 2 _CP4_SEARCH_PROBE_IMPROVEMENT_MS = 1.0 _PLAN_CACHE_MAX_ENTRIES = 128 @@ -181,6 +181,50 @@ def _get_or_build_planning_bundle( return planning_key, bundle, group_ids_cpu, parent_ids_cpu +@dataclass(frozen=True) +class ContextParallelPlanSummary: + """Structure of one packed row's context-parallel plan: the number of + remote attention waves and the largest per-rank token load. Derived from + the cached planning bundle, so a caller that summarizes a row and then + executes it pays for the assignment search once.""" + + wave_count: int + max_rank_tokens: int + + +def summarize_prefix_tree_plan( + *, + group_ids: torch.Tensor, + parent_ids: torch.Tensor, + topology: ParallelTopology, + config: ContextParallelConfig, + original_seq_len: int, + build_gdn_execution_spec: bool = False, +) -> ContextParallelPlanSummary: + """Plan one packed row (CPU metadata) and report its CP plan structure. + + The layout planner's second stage prices shortlisted layouts with this; + the bundle it builds is the one ``prepare_cp_micro`` reuses for the layout + that is finally executed. + """ + + _planning_key, bundle, _group_ids, _parent_ids = _get_or_build_planning_bundle( + group_ids=group_ids, + parent_ids=parent_ids, + topology=topology, + config=config, + original_seq_len=original_seq_len, + build_gdn_execution_spec=build_gdn_execution_spec, + ) + wave_count = max(bundle.wave_assignment) + 1 if bundle.wave_assignment else 0 + return ContextParallelPlanSummary( + wave_count=int(wave_count), + max_rank_tokens=int( + max(bundle.token_layout_index.token_counts_by_rank, default=0) + ), + ) + + def _get_or_build_bundle_rank_plan( *, planning_key: str, @@ -472,7 +516,7 @@ def _best_improving_move( cp_size: int, q_weights: list[float], candidate_limit: int, - evaluate_candidate: Any, + evaluate_candidates: Any, ) -> tuple[tuple[int, ...], dict[str, Any]] | None: slow_rank = int( max( @@ -489,7 +533,7 @@ def _best_improving_move( if not candidate_chunks: return None - best_move: tuple[tuple[int, ...], dict[str, Any]] | None = None + moves: list[tuple[int, ...]] = [] for chunk_index in candidate_chunks: for dst_rank in range(cp_size): if dst_rank == slow_rank: @@ -502,16 +546,19 @@ def _best_improving_move( and len(set(candidate_owners)) != cp_size ): continue - candidate_eval = evaluate_candidate( - owners=candidate_owners, - wave_assignment=wave_assignment, - ) - if float(candidate_eval["score"]) + 1e-9 >= float(current_eval["score"]): - continue - if best_move is None or float(candidate_eval["score"]) + 1e-9 < float( - best_move[1]["score"] - ): - best_move = (candidate_owners, candidate_eval) + moves.append(candidate_owners) + evaluations = evaluate_candidates( + owners_list=moves, + wave_assignment=wave_assignment, + ) + best_move: tuple[tuple[int, ...], dict[str, Any]] | None = None + for candidate_owners, candidate_eval in zip(moves, evaluations, strict=True): + if float(candidate_eval["score"]) + 1e-9 >= float(current_eval["score"]): + continue + if best_move is None or float(candidate_eval["score"]) + 1e-9 < float( + best_move[1]["score"] + ): + best_move = (candidate_owners, candidate_eval) return best_move @@ -910,37 +957,55 @@ def _ranges_size(ranges: tuple[TokenRange, ...]) -> int: return int(sum(range_.size() for range_ in ranges)) -def _chunk_mask_stats( +@dataclass(frozen=True) +class _PairProgram: + """The chunk pair program as the plan evaluator consumes it. + + Pair counts and their positivity are float64 matrices: every value is an + integer far below 2**53, so the BLAS-backed products below are exact, and + the evaluator only ever compares them with zero or converts them back to + ``int``. + """ + + pair_counts: np.ndarray + pair_positive: np.ndarray + chunk_lengths: np.ndarray + + +def _pair_program( + pair_matrix: list[list[int]] | torch.Tensor, *, - chunk_lengths: tuple[int, ...], - chunk_mask: torch.Tensor, - chunk_lengths_tensor: torch.Tensor | None = None, -) -> tuple[int, int]: - if ( - chunk_lengths_tensor is not None - and len(chunk_lengths) >= _CHUNK_MASK_STATS_TORCH_THRESHOLD - ): - if int(chunk_mask.numel()) == 0 or not bool(chunk_mask.any().item()): - return 0, 0 - token_count = int(chunk_lengths_tensor[chunk_mask].sum().item()) - run_starts = chunk_mask.clone() - run_starts[1:] = torch.logical_and( - run_starts[1:], torch.logical_not(chunk_mask[:-1]) - ) - range_count = int(run_starts.sum().item()) - return token_count, range_count - token_count = 0 - range_count = 0 - in_run = False - for is_set, length in zip(chunk_mask.tolist(), chunk_lengths, strict=True): - if bool(is_set): - token_count += int(length) - if not in_run: - range_count += 1 - in_run = True - continue - in_run = False - return token_count, range_count + chunk_ranges: tuple[TokenRange, ...], +) -> _PairProgram: + chunk_count = len(chunk_ranges) + counts = ( + pair_matrix.to(torch.float64).numpy() + if isinstance(pair_matrix, torch.Tensor) + else np.asarray(pair_matrix, dtype=np.float64) + ).reshape(chunk_count, chunk_count) + return _PairProgram( + pair_counts=counts, + pair_positive=(counts > 0).astype(np.float64), + chunk_lengths=np.asarray( + [int(range_.size()) for range_ in chunk_ranges], dtype=np.int64 + ), + ) + + +def _mask_token_and_range_counts( + masks: np.ndarray, + chunk_lengths: np.ndarray, +) -> tuple[np.ndarray, np.ndarray]: + """For every chunk mask along the last axis: the tokens it covers and the + number of contiguous chunk runs it consists of.""" + + tokens = masks.astype(np.int64) @ chunk_lengths + if masks.shape[-1] == 0: + return tokens, np.zeros(masks.shape[:-1], dtype=np.int64) + runs = masks[..., 0].astype(np.int64) + (masks[..., 1:] & ~masks[..., :-1]).sum( + axis=-1, dtype=np.int64 + ) + return tokens, runs def _stage_cost_ms( @@ -1057,267 +1122,292 @@ def _simulate_backward_time_ms( return max(current_time, max(reduce_ready_times, default=0.0)) -def _evaluate_plan( +def _stage_cost_ms_array( *, - chunk_ranges: tuple[TokenRange, ...], - pair_matrix: list[list[int]] | torch.Tensor, - owners: tuple[int, ...], + pair_count: np.ndarray, + q_tokens: np.ndarray, + k_tokens: np.ndarray, + q_range_count: np.ndarray, + k_range_count: np.ndarray, + config: ContextParallelConfig, + backward: bool, + local: bool, +) -> np.ndarray: + """``_stage_cost_ms`` over arrays of integer-valued float64 counts, term by + term in the same order, so every element equals the scalar formula.""" + + pair_ms = ( + config.planner_local_backward_pair_ms + if backward and local + else ( + config.planner_remote_backward_pair_ms + if backward + else ( + config.planner_local_pair_ms if local else config.planner_remote_pair_ms + ) + ) + ) + remote_underfill_ms = np.zeros_like(pair_count) + if not local: + token_floor = int(config.planner_remote_stage_token_floor) + pair_floor = int(config.planner_remote_stage_pair_floor) + token_shortfall = np.maximum(token_floor - np.minimum(q_tokens, k_tokens), 0.0) + pair_shortfall = np.maximum(pair_floor - pair_count, 0.0) + token_scale = ( + token_shortfall / float(token_floor) + if token_floor > 0 + else np.zeros_like(pair_count) + ) + pair_scale = ( + pair_shortfall / float(pair_floor) + if pair_floor > 0 + else np.zeros_like(pair_count) + ) + remote_underfill_ms = np.where( + (pair_count > 0) | (q_tokens > 0) | (k_tokens > 0), + float(config.planner_remote_stage_underfill_ms) + * np.maximum(token_scale, pair_scale), + 0.0, + ) + return ( + float(config.planner_stage_overhead_ms) + + pair_count * float(pair_ms) + + q_tokens * float(config.planner_merge_q_token_ms) + + (q_range_count + k_range_count) * float(config.planner_interval_overhead_ms) + + remote_underfill_ms + ) + + +def _comm_cost_ms_array( + *, + tokens: np.ndarray, + range_count: np.ndarray, + config: ContextParallelConfig, + backward: bool, +) -> np.ndarray: + per_token = ( + float(config.planner_reduce_token_ms) + if backward + else float(config.planner_fetch_token_ms) + ) + return np.where( + (tokens <= 0) & (range_count <= 0), + 0.0, + float(config.planner_comm_stage_overhead_ms) + + tokens * per_token + + range_count * float(config.planner_interval_overhead_ms), + ) + + +def _simulate_forward_time_ms_array( + *, + local_stage_ms: np.ndarray, + remote_stage_ms: np.ndarray, + remote_fetch_ms: np.ndarray, + active: np.ndarray, +) -> np.ndarray: + """``_simulate_forward_time_ms`` over (..., wave) arrays whose inactive + waves are skipped exactly as the scalar version skips absent stages.""" + + wave_count = active.shape[-1] + # The fetch of the next active wave after each wave (a reverse scan). + next_fetch = np.zeros_like(remote_fetch_ms) + has_next = np.zeros(active.shape, dtype=bool) + carry_fetch = np.zeros_like(local_stage_ms) + carry_has = np.zeros(local_stage_ms.shape, dtype=bool) + for wave in reversed(range(wave_count)): + next_fetch[..., wave] = carry_fetch + has_next[..., wave] = carry_has + carry_fetch = np.where( + active[..., wave], remote_fetch_ms[..., wave], carry_fetch + ) + carry_has = carry_has | active[..., wave] + current = local_stage_ms.copy() + fetch_ready = np.zeros_like(local_stage_ms) + started = np.zeros(local_stage_ms.shape, dtype=bool) + for wave in range(wave_count): + is_active = active[..., wave] + ready = np.where(started, fetch_ready, remote_fetch_ms[..., wave]) + compute_start = np.maximum(current, ready) + current = np.where( + is_active, compute_start + remote_stage_ms[..., wave], current + ) + fetch_ready = np.where( + is_active & has_next[..., wave], + compute_start + next_fetch[..., wave], + fetch_ready, + ) + started = started | is_active + return current + + +def _simulate_backward_time_ms_array( + *, + local_stage_ms: np.ndarray, + remote_stage_ms: np.ndarray, + remote_reduce_ms: np.ndarray, + active: np.ndarray, +) -> np.ndarray: + current = np.zeros_like(local_stage_ms) + reduce_ready = np.zeros_like(local_stage_ms) + for wave in range(active.shape[-1]): + is_active = active[..., wave] + current = np.where(is_active, current + remote_stage_ms[..., wave], current) + reduce_ready = np.where( + is_active, + np.maximum(reduce_ready, current + remote_reduce_ms[..., wave]), + reduce_ready, + ) + total = current + local_stage_ms + return np.where( + active.any(axis=-1), np.maximum(total, reduce_ready), local_stage_ms + ) + + +def _evaluate_plans( + *, + program: _PairProgram, + owners_batch: np.ndarray, wave_assignment: tuple[int, ...], cp_size: int, config: ContextParallelConfig, - pair_positive: torch.Tensor | None = None, - chunk_lengths: tuple[int, ...] | None = None, - chunk_lengths_tensor: torch.Tensor | None = None, -) -> dict[str, Any]: - rank_scores: list[float] = [] - rank_forward_ms: list[float] = [] - rank_backward_ms: list[float] = [] - chunk_count = len(chunk_ranges) +) -> list[dict[str, Any]]: + """Predicted per-rank forward and backward time of a batch of chunk + assignments (``owners_batch``: candidate x chunk) under one wave assignment. + + The mask algebra (which chunks each rank owns, attends, fetches and sends + per wave), the cost formulas and the stage simulations are all vectorized + over candidates, ranks, sources and waves; each candidate prices exactly as + it would on its own. + """ + + batch = int(owners_batch.shape[0]) + chunk_count = int(program.chunk_lengths.shape[0]) wave_count = max(wave_assignment, default=0) + 1 if wave_assignment else 0 - pair_counts = ( - pair_matrix - if isinstance(pair_matrix, torch.Tensor) and pair_matrix.dtype == torch.int64 - else torch.as_tensor(pair_matrix, dtype=torch.int64) - ) - if pair_positive is None: - pair_positive = pair_counts > 0 - if chunk_lengths is None: - chunk_lengths = tuple(int(range_.size()) for range_ in chunk_ranges) - if ( - chunk_lengths_tensor is None - and len(chunk_lengths) >= _CHUNK_MASK_STATS_TORCH_THRESHOLD - ): - chunk_lengths_tensor = torch.tensor(chunk_lengths, dtype=torch.int64) - owners_tensor = torch.tensor(owners, dtype=torch.int64) - wave_tensor = torch.tensor( - wave_assignment, - dtype=torch.int64, - ) - owner_masks = [owners_tensor == rank for rank in range(cp_size)] - owner_indices = [ - torch.nonzero(owner_mask, as_tuple=False).flatten() - for owner_mask in owner_masks - ] - empty_pair_counts = pair_counts.new_zeros((0, chunk_count)) - empty_pair_positive = pair_positive.new_zeros((0, chunk_count)) - pair_counts_by_rank_rows = [ - ( - empty_pair_counts - if int(owner_index.numel()) == 0 - else pair_counts.index_select(0, owner_index) - ) - for owner_index in owner_indices - ] - pair_positive_by_rank_rows = [ - ( - empty_pair_positive - if int(owner_index.numel()) == 0 - else pair_positive.index_select(0, owner_index) - ) - for owner_index in owner_indices - ] - pair_positive_by_rank_cols = [ - ( - torch.zeros(chunk_count, dtype=torch.bool) - if int(rank_rows.numel()) == 0 - else rank_rows.any(dim=0) - ) - for rank_rows in pair_positive_by_rank_rows + lengths = program.chunk_lengths + waves_array = np.asarray(wave_assignment, dtype=np.int64).reshape(chunk_count) + # (candidate, rank, chunk): ownership; (wave, chunk): wave membership. + owned = owners_batch[:, None, :] == np.arange(cp_size)[None, :, None] + in_wave = waves_array[None, :] == np.arange(wave_count)[:, None] + owned_f = owned.astype(np.float64) + # (candidate, rank, k chunk): pairs between the rank's q chunks and each k chunk. + owned_pair_counts = owned_f @ program.pair_counts + positive_cols = owned_pair_counts > 0 + local_pairs = (owned_pair_counts * owned_f).sum(axis=2) + local_q = owned & ((owned_f @ program.pair_positive.T) > 0) + local_k = owned & positive_cols + local_q_tokens, local_q_ranges = _mask_token_and_range_counts(local_q, lengths) + local_k_tokens, local_k_ranges = _mask_token_and_range_counts(local_k, lengths) + + not_self = ~np.eye(cp_size, dtype=bool) + # (candidate, rank, source, wave, chunk): the source's chunks in this wave + # that the rank's queries attend, i.e. what the rank fetches from it. + touched_source = ( + owned[:, None, :, None, :] + & in_wave[None, None, None, :, :] + & positive_cols[:, :, None, None, :] + & not_self[None, :, :, None, None] + ) + request_tokens, request_ranges = _mask_token_and_range_counts( + touched_source, lengths + ) + recv_tokens = request_tokens.sum(axis=2) + recv_ranges = request_ranges.sum(axis=2) + touched = touched_source.any(axis=2) # (candidate, rank, wave, chunk) + touched_f = touched.astype(np.float64) + request_pairs = np.einsum("crj,crwj->crw", owned_pair_counts, touched_f) + touched_q = owned[:, :, None, :] & ((touched_f @ program.pair_positive.T) > 0) + q_tokens, q_ranges = _mask_token_and_range_counts(touched_q, lengths) + # (candidate, rank, wave, peer, chunk): the rank's owned chunks in this + # wave that the peer's queries attend, i.e. what the rank sends to it. + send_mask = ( + (owned[:, :, None, :] & in_wave[None, None, :, :])[:, :, :, None, :] + & positive_cols[:, None, None, :, :] + & not_self[None, :, None, :, None] + ) + send_tokens_by_peer, send_ranges_by_peer = _mask_token_and_range_counts( + send_mask, lengths + ) + aggregate_tokens, aggregate_ranges = _mask_token_and_range_counts( + send_mask.any(axis=3), lengths + ) + send_tokens = send_tokens_by_peer.sum(axis=3) + aggregate_tokens + send_ranges = send_ranges_by_peer.sum(axis=3) + aggregate_ranges + active = (request_pairs > 0) | (recv_tokens > 0) | (recv_ranges > 0) + + local_costs: dict[str, Any] = dict( + pair_count=local_pairs, + q_tokens=local_q_tokens.astype(np.float64), + k_tokens=local_k_tokens.astype(np.float64), + q_range_count=local_q_ranges.astype(np.float64), + k_range_count=local_k_ranges.astype(np.float64), + config=config, + local=True, + ) + local_forward = np.where( + local_pairs > 0, _stage_cost_ms_array(backward=False, **local_costs), 0.0 + ) + local_backward = np.where( + local_pairs > 0, _stage_cost_ms_array(backward=True, **local_costs), 0.0 + ) + stage_costs: dict[str, Any] = dict( + pair_count=request_pairs, + q_tokens=q_tokens.astype(np.float64), + k_tokens=recv_tokens.astype(np.float64), + q_range_count=q_ranges.astype(np.float64), + k_range_count=recv_ranges.astype(np.float64), + config=config, + local=False, + ) + comm_costs: dict[str, Any] = dict( + tokens=np.maximum(send_tokens, recv_tokens).astype(np.float64), + range_count=np.maximum(send_ranges, recv_ranges).astype(np.float64), + config=config, + ) + forward_ms = _simulate_forward_time_ms_array( + local_stage_ms=local_forward, + remote_stage_ms=_stage_cost_ms_array(backward=False, **stage_costs), + remote_fetch_ms=_comm_cost_ms_array(backward=False, **comm_costs), + active=active, + ) + backward_ms = _simulate_backward_time_ms_array( + local_stage_ms=local_backward, + remote_stage_ms=_stage_cost_ms_array(backward=True, **stage_costs), + remote_reduce_ms=_comm_cost_ms_array(backward=True, **comm_costs), + active=active, + ) + rank_scores = forward_ms + backward_ms + return [ + { + "score": float(rank_scores[index].max()) if cp_size else 0.0, + "rank_scores": tuple(float(value) for value in rank_scores[index]), + "rank_forward_ms": tuple(float(value) for value in forward_ms[index]), + "rank_backward_ms": tuple(float(value) for value in backward_ms[index]), + } + for index in range(batch) ] - wave_masks = [wave_tensor == wave_index for wave_index in range(wave_count)] - - for rank in range(cp_size): - owned_q_mask = owner_masks[rank] - owned_q_indices = owner_indices[rank] - owned_pair_counts = pair_counts_by_rank_rows[rank] - owned_pair_positive = pair_positive_by_rank_rows[rank] - owned_positive_cols = pair_positive_by_rank_cols[rank] - - local_pairs = ( - 0 - if int(owned_q_indices.numel()) == 0 - else int(owned_pair_counts.index_select(1, owned_q_indices).sum().item()) - ) - local_q_mask = torch.zeros(chunk_count, dtype=torch.bool) - if int(owned_q_indices.numel()) > 0: - touched_local_q = owned_pair_positive.index_select(1, owned_q_indices).any( - dim=1 - ) - if bool(touched_local_q.any().item()): - local_q_mask[owned_q_indices[touched_local_q]] = True - local_k_mask = owned_q_mask & owned_positive_cols - local_q_tokens, local_q_range_count = _chunk_mask_stats( - chunk_lengths=chunk_lengths, - chunk_mask=local_q_mask, - chunk_lengths_tensor=chunk_lengths_tensor, - ) - local_k_tokens, local_k_range_count = _chunk_mask_stats( - chunk_lengths=chunk_lengths, - chunk_mask=local_k_mask, - chunk_lengths_tensor=chunk_lengths_tensor, - ) - local_stage_ms = _stage_cost_ms( - pair_count=local_pairs, - q_tokens=local_q_tokens, - k_tokens=local_k_tokens, - q_range_count=local_q_range_count, - k_range_count=local_k_range_count, - config=config, - backward=False, - local=True, - ) - local_backward_ms = _stage_cost_ms( - pair_count=local_pairs, - q_tokens=local_q_tokens, - k_tokens=local_k_tokens, - q_range_count=local_q_range_count, - k_range_count=local_k_range_count, - config=config, - backward=True, - local=True, - ) - - remote_stage_ms: list[float] = [] - remote_fetch_ms: list[float] = [] - remote_backward_ms: list[float] = [] - remote_reduce_ms: list[float] = [] - for wave_index in range(wave_count): - request_tokens_by_source = [0 for _ in range(cp_size)] - request_range_counts_by_source = [0 for _ in range(cp_size)] - request_pairs = 0 - touched_q_mask = torch.zeros(chunk_count, dtype=torch.bool) - for source_rank in range(cp_size): - if source_rank == rank: - continue - touched_source_mask = ( - owner_masks[source_rank] - & wave_masks[wave_index] - & owned_positive_cols - ) - ( - request_tokens_by_source[source_rank], - request_range_counts_by_source[source_rank], - ) = _chunk_mask_stats( - chunk_lengths=chunk_lengths, - chunk_mask=touched_source_mask, - chunk_lengths_tensor=chunk_lengths_tensor, - ) - if request_tokens_by_source[source_rank] <= 0: - continue - touched_source_indices = torch.nonzero( - touched_source_mask, - as_tuple=False, - ).flatten() - request_pairs += int( - owned_pair_counts.index_select(1, touched_source_indices) - .sum() - .item() - ) - touched_remote_q = owned_pair_positive.index_select( - 1, - touched_source_indices, - ).any(dim=1) - if bool(touched_remote_q.any().item()): - touched_q_mask[owned_q_indices[touched_remote_q]] = True - recv_tokens = sum(request_tokens_by_source) - recv_range_count = sum(request_range_counts_by_source) - if request_pairs <= 0 and recv_tokens <= 0 and recv_range_count <= 0: - continue - send_tokens_by_peer = [0 for _ in range(cp_size)] - send_range_counts_by_peer = [0 for _ in range(cp_size)] - aggregate_send_mask = torch.zeros(chunk_count, dtype=torch.bool) - owned_wave_mask = owned_q_mask & wave_masks[wave_index] - if bool(owned_wave_mask.any().item()): - for peer_rank in range(cp_size): - if peer_rank == rank: - continue - send_mask = owned_wave_mask & pair_positive_by_rank_cols[peer_rank] - ( - send_tokens_by_peer[peer_rank], - send_range_counts_by_peer[peer_rank], - ) = _chunk_mask_stats( - chunk_lengths=chunk_lengths, - chunk_mask=send_mask, - chunk_lengths_tensor=chunk_lengths_tensor, - ) - if send_tokens_by_peer[peer_rank] > 0: - aggregate_send_mask |= send_mask - ( - send_tokens_by_peer[rank], - send_range_counts_by_peer[rank], - ) = _chunk_mask_stats( - chunk_lengths=chunk_lengths, - chunk_mask=aggregate_send_mask, - chunk_lengths_tensor=chunk_lengths_tensor, - ) - send_tokens = sum(send_tokens_by_peer) - q_tokens, q_range_count = _chunk_mask_stats( - chunk_lengths=chunk_lengths, - chunk_mask=touched_q_mask, - chunk_lengths_tensor=chunk_lengths_tensor, - ) - remote_stage_ms.append( - _stage_cost_ms( - pair_count=request_pairs, - q_tokens=q_tokens, - k_tokens=recv_tokens, - q_range_count=q_range_count, - k_range_count=recv_range_count, - config=config, - backward=False, - local=False, - ) - ) - remote_backward_ms.append( - _stage_cost_ms( - pair_count=request_pairs, - q_tokens=q_tokens, - k_tokens=recv_tokens, - q_range_count=q_range_count, - k_range_count=recv_range_count, - config=config, - backward=True, - local=False, - ) - ) - remote_fetch_ms.append( - _comm_cost_ms( - tokens=max(send_tokens, recv_tokens), - range_count=max(sum(send_range_counts_by_peer), recv_range_count), - config=config, - backward=False, - ) - ) - remote_reduce_ms.append( - _comm_cost_ms( - tokens=max(send_tokens, recv_tokens), - range_count=max(sum(send_range_counts_by_peer), recv_range_count), - config=config, - backward=True, - ) - ) +def _evaluate_plan( + *, + program: _PairProgram, + owners: tuple[int, ...], + wave_assignment: tuple[int, ...], + cp_size: int, + config: ContextParallelConfig, +) -> dict[str, Any]: + """Predicted per-rank forward and backward time of one chunk assignment.""" - forward_ms = _simulate_forward_time_ms( - local_stage_ms=local_stage_ms if local_pairs > 0 else 0.0, - remote_stage_ms=tuple(remote_stage_ms), - remote_fetch_ms=tuple(remote_fetch_ms), - ) - backward_ms = _simulate_backward_time_ms( - local_stage_ms=local_backward_ms if local_pairs > 0 else 0.0, - remote_stage_ms=tuple(remote_backward_ms), - remote_reduce_ms=tuple(remote_reduce_ms), - ) - rank_forward_ms.append(float(forward_ms)) - rank_backward_ms.append(float(backward_ms)) - rank_scores.append(float(forward_ms + backward_ms)) - return { - "score": max(rank_scores, default=0.0), - "rank_scores": tuple(rank_scores), - "rank_forward_ms": tuple(rank_forward_ms), - "rank_backward_ms": tuple(rank_backward_ms), - } + owners_batch = np.asarray(owners, dtype=np.int64).reshape( + 1, int(program.chunk_lengths.shape[0]) + ) + return _evaluate_plans( + program=program, + owners_batch=owners_batch, + wave_assignment=wave_assignment, + cp_size=cp_size, + config=config, + )[0] def _search_generic_chunk_assignment( @@ -1338,37 +1428,41 @@ def _search_generic_chunk_assignment( ) best: tuple[tuple[int, ...], tuple[int, ...], dict[str, Any]] | None = None eval_cache: dict[tuple[tuple[int, ...], tuple[int, ...]], dict[str, Any]] = {} - pair_counts = torch.as_tensor(pair_matrix, dtype=torch.int64) - pair_positive = pair_counts > 0 - chunk_lengths = tuple(int(range_.size()) for range_ in chunk_ranges) - chunk_lengths_tensor = ( - torch.tensor(chunk_lengths, dtype=torch.int64) - if len(chunk_lengths) >= _CHUNK_MASK_STATS_TORCH_THRESHOLD - else None - ) + program = _pair_program(pair_matrix, chunk_ranges=chunk_ranges) + + def _evaluate_candidates( + *, + owners_list: list[tuple[int, ...]], + wave_assignment: tuple[int, ...], + ) -> list[dict[str, Any]]: + misses = [ + owners + for owners in dict.fromkeys(owners_list) + if (owners, wave_assignment) not in eval_cache + ] + if misses: + evaluations = _evaluate_plans( + program=program, + owners_batch=np.asarray(misses, dtype=np.int64).reshape( + len(misses), len(chunk_ranges) + ), + wave_assignment=wave_assignment, + cp_size=cp_size, + config=config, + ) + for owners, evaluation in zip(misses, evaluations, strict=True): + eval_cache[(owners, wave_assignment)] = evaluation + return [eval_cache[(owners, wave_assignment)] for owners in owners_list] def _evaluate_candidate( *, owners: tuple[int, ...], wave_assignment: tuple[int, ...], ) -> dict[str, Any]: - cache_key = (owners, wave_assignment) - cached = eval_cache.get(cache_key) - if cached is not None: - return cached - cached = _evaluate_plan( - chunk_ranges=chunk_ranges, - pair_matrix=pair_counts, - owners=owners, + return _evaluate_candidates( + owners_list=[owners], wave_assignment=wave_assignment, - cp_size=cp_size, - config=config, - pair_positive=pair_positive, - chunk_lengths=chunk_lengths, - chunk_lengths_tensor=chunk_lengths_tensor, - ) - eval_cache[cache_key] = cached - return cached + )[0] contiguous_owners = _contiguous_chunk_assignment( q_weights=q_weights, @@ -1410,7 +1504,7 @@ def _evaluate_candidate( int(config.planner_candidate_chunk_limit), _CP4_SEARCH_PROBE_CANDIDATE_LIMIT, ), - evaluate_candidate=_evaluate_candidate, + evaluate_candidates=_evaluate_candidates, ) if ( probe_move is not None @@ -1430,7 +1524,7 @@ def _evaluate_candidate( cp_size=cp_size, q_weights=q_weights, candidate_limit=int(config.planner_candidate_chunk_limit), - evaluate_candidate=_evaluate_candidate, + evaluate_candidates=_evaluate_candidates, ) if best_move is None: break diff --git a/src/art/trainer_rank/_impl.py b/src/art/trainer_rank/_impl.py index bdde05797..4665e6a98 100644 --- a/src/art/trainer_rank/_impl.py +++ b/src/art/trainer_rank/_impl.py @@ -1207,6 +1207,8 @@ def __init__(self, runtime: TrainingRuntime) -> None: self._coefficient_version = selection.version self._coefficient_table = selection.table_id self._planner_coefficients = selection.coefficients + # Two-stage selection on re-ranked shapes (part of the table identity). + self._planner_reranker = selection.reranker if selection.table_id is None: logger.warning( "TrainerRank layout cost model: runtime (capability=%s, device " @@ -3746,6 +3748,14 @@ def _compute_group_layout( coefficient_version=facts.coefficient_version, coefficients=self._planner_coefficients, refinement_work_budget=_PLANNER_REFINEMENT_BUDGET, + reranker=self._planner_reranker, + plan_structure=( + None + if self._planner_reranker is None + else lambda candidate: self._plan_structure( + input_ids, tree, candidate + ) + ), ).layout cached = (tree, layout) with self._layout_cache_lock: @@ -3755,6 +3765,39 @@ def _compute_group_layout( self._layout_selection_cache.popitem(last=False) return cached + def _plan_structure( + self, + input_ids: Sequence[torch.Tensor], + tree: CanonicalPrefixTree, + layout: PrefixTreeLayout, + ) -> tuple[int, int]: + """(remote wave count, largest per-rank token load) of the context- + parallel plan this layout produces, through the runtime's planning + bundle cache: the layout finally selected reuses its bundle.""" + + from art.megatron.context_parallel.runtime import summarize_prefix_tree_plan + from art.megatron.training.microbatches import ( + _context_parallel_config_for_provider, + ) + + packed = materialize_prefix_tree_layout( + input_ids, tree, layout, verify_shared_tokens=False + ) + handler = self.runtime.model_support_handler + summary = summarize_prefix_tree_plan( + group_ids=packed.group_ids, + parent_ids=packed.parent_ids, + topology=self._topology(), + config=_context_parallel_config_for_provider( + self.runtime.provider, self.device, handler + ), + original_seq_len=int(packed.tokens.shape[1]), + build_gdn_execution_spec=bool( + getattr(handler, "build_gdn_execution_spec", False) + ), + ) + return summary.wave_count, summary.max_rank_tokens + def _select_group_layout( self, input_ids: Sequence[torch.Tensor], @@ -5017,7 +5060,14 @@ def _configure_hybridep( rows = tuple(self._hybridep_rows(batch, topology=topology) for batch in padded) current = fused_a2a._hybrid_ep_buffer live = self._has_live_hybridep_graphs() - required_capacity = _hybridep_token_capacity(sequence_length, int(topology.cp)) + # The buffer must hold the busiest rank's planned rows: cost-aware CP + # plans (few long histories, shared prefixes) can put more than the + # near-balanced extent on one rank, and every rank sizes the buffer + # identically because the planned counts are the same on all ranks. + required_capacity = max( + _hybridep_token_capacity(sequence_length, int(topology.cp)), + max(rows), + ) if live and ( current is None or id(current) != getattr(self, "_hybridep_buffer_id", None) @@ -5033,6 +5083,7 @@ def _configure_hybridep( self.runtime, packed_sequence_length=sequence_length, context_parallel_size=int(topology.cp), + required_capacity=required_capacity, ) current = fused_a2a._hybrid_ep_buffer if current is None: diff --git a/src/art/trainer_rank/_planner_cost.py b/src/art/trainer_rank/_planner_cost.py index 0a050ff4c..a66a9375f 100644 --- a/src/art/trainer_rank/_planner_cost.py +++ b/src/art/trainer_rank/_planner_cost.py @@ -15,7 +15,7 @@ deltas, refined by direct regret minimization, fitted on 45 cells and evaluated on all 56 (the 11 odd Ellavox groups are the pre-registered holdout; whole topologies, shapes and models were withheld in ablations). The table is -bound to its evidence by ``dev/trainer_rank_cost_calibration_certificate.json``. +bound to its evidence by ``dev/trainer_rank_cost_calibration_certificate_
.json``. The version-1 constants they replace were hand-set, not fitted. What the data showed: the cost of a shared prefix level is a GDN effect that @@ -327,6 +327,34 @@ class ParallelShape: etp: int = 1 +@dataclass(frozen=True) +class ReRanker: + """Second-stage score for a table's re-ranked parallel shapes. + + On these shapes the ten-term score cannot see the structure of the + context-parallel plan a layout produces, which dominates real-data cells + at CP4. It still shortlists well: the ``shortlist_size`` cheapest layouts + of the search under ``shortlist_coefficients_milli_us`` (plus the + ``incumbent`` anchor) are then priced per layer by the plan they produce, + remote wave count and largest per-rank token load, and the lowest + second-stage score is selected. Fitted and certified per class and shape + like the tables (shortlist recall, ranking gates, non-regression against + the cheap score and version 1, planning cost). + """ + + shortlist_size: int + incumbent: str + shortlist_coefficients_milli_us: Mapping[str, int] + wave_per_layer_milli_us: int + max_rank_token_per_layer_milli_us: int + + def score(self, *, layers: int, wave_count: int, max_rank_tokens: int) -> int: + return max(1, layers) * ( + wave_count * self.wave_per_layer_milli_us + + max_rank_tokens * self.max_rank_token_per_layer_milli_us + ) + + @dataclass(frozen=True) class CalibratedTable: """One fitted coefficient table and the execution classes it admits. @@ -335,6 +363,8 @@ class CalibratedTable: geometry and parallel shape must each appear in the table's measured sets (the certificate test binds these sets to the recorded cells). No configuration is admitted because a width falls between measured widths. + ``reranked_shapes`` are admitted through the two-stage selection of the + table's ``reranker`` instead of the ten-term score alone. """ table_id: str @@ -343,30 +373,72 @@ class CalibratedTable: param_dtypes: tuple[str, ...] geometries: tuple[ModelGeometry, ...] shapes: tuple[ParallelShape, ...] - - def admits( + # Measured (geometry, shape) pairs the table does not admit: their cells + # are in the certificate and the fit, but the score is known to misrank + # that pair (a documented blind spot), so the fallback applies there. + withheld: tuple[tuple[ModelGeometry, ParallelShape], ...] = () + reranked_shapes: tuple[ParallelShape, ...] = () + reranker: ReRanker | None = None + + def __post_init__(self) -> None: + if set(self.reranked_shapes) & set(self.shapes): + raise ValueError("a shape is either scored directly or re-ranked") + if bool(self.reranked_shapes) != (self.reranker is not None): + raise ValueError("re-ranked shapes and a re-ranker come together") + + def admits_class( self, *, device: DeviceClass, param_dtype: str, geometry: ModelGeometry, - shape: ParallelShape, ) -> bool: return ( any(known.admits(device) for known in self.device_classes) and param_dtype in self.param_dtypes and geometry in self.geometries + ) + + def admits( + self, + *, + device: DeviceClass, + param_dtype: str, + geometry: ModelGeometry, + shape: ParallelShape, + ) -> bool: + return ( + self.admits_class(device=device, param_dtype=param_dtype, geometry=geometry) and shape in self.shapes + and (geometry, shape) not in self.withheld + ) + + def reranks( + self, + *, + device: DeviceClass, + param_dtype: str, + geometry: ModelGeometry, + shape: ParallelShape, + ) -> bool: + return ( + self.reranker is not None + and self.admits_class( + device=device, param_dtype=param_dtype, geometry=geometry + ) + and shape in self.reranked_shapes ) @dataclass(frozen=True, slots=True) class ScoringSelection: - """The score a runtime uses: version 2 with one table, or the fallback.""" + """The score a runtime uses: version 2 with one table (directly, or as + the shortlist score of a two-stage re-ranker), or the fallback.""" version: int table_id: str | None coefficients: Mapping[str, int] | None + reranker: ReRanker | None = None def select_scoring( @@ -399,6 +471,16 @@ def select_scoring( return ScoringSelection( COEFFICIENT_VERSION, table.table_id, table.coefficients_milli_us ) + if table.reranks( + device=device, param_dtype=param_dtype, geometry=geometry, shape=shape + ): + assert table.reranker is not None + return ScoringSelection( + COEFFICIENT_VERSION, + table.table_id, + table.reranker.shortlist_coefficients_milli_us, + table.reranker, + ) return ScoringSelection(COEFFICIENT_VERSION_FALLBACK, None, None) @@ -497,11 +579,334 @@ def _prefix_tree_layout_score_v1( ParallelShape(tp=1, cp=4), ParallelShape(tp=2, cp=1), ), + # Qwen3-4B at TP1 x CP4 is measured but withheld: on real-data groups the + # ten-term score cannot see the context-parallel exchange schedule and its + # regret reached 35% (design brief, known limitation); version 1 is the + # less harmful fallback there until the CP-aware re-ranker is certified. + withheld=((QWEN3_4B_GEOMETRY, ParallelShape(tp=1, cp=4)),), +) +# Qwen3.5-35B-A3B class (GDN + MoE, hidden 2,048): 16 attention heads in 2 query groups of 256 channels, GDN 32 value / 16 key heads of 128 dims, 256 experts (top-8, expert FFN 512, shared expert 512), on H200 bf16; measured at TP1 x CP1/2/4 and TP2 with EP1, EP2 at CP2/CP4/TP2 and EP4 at CP4 (2026-09-04). +GDN_MOE_H2048_GEOMETRY = ModelGeometry( + ffn_hidden_size=8_192, + gdn_conv_kernel=4, + gdn_key_head_dim=128, + gdn_key_heads=16, + gdn_value_head_dim=128, + gdn_value_heads=32, + hidden_size=2_048, + kv_channels=256, + moe_experts=256, + moe_ffn_hidden_size=512, + moe_shared_expert_ffn=512, + moe_topk=8, + num_attention_heads=16, + num_query_groups=2, +) +GDN_MOE_H2048_TABLE = CalibratedTable( + table_id="gdn-moe-h2048-h200-bf16", + coefficients_milli_us={ + "attention_token_cp_exchange": 37, + "gdn_level": 8_982_505, + "gdn_level_tp": 0, + "gdn_token_per_rank": 368, + "level_cp_per_layer": 50_356, + "level_tp_per_layer": 5_701_857, + "tiny_segment_per_layer": 442_444, + "token_cp_exchange": 37, + "token_per_rank": 3_145, + "token_tp_collective": 734, + }, + device_classes=(H200_CLASS,), + param_dtypes=("torch.bfloat16",), + geometries=(GDN_MOE_H2048_GEOMETRY,), + shapes=( + ParallelShape(tp=1, cp=1), + ParallelShape(tp=1, cp=2), + ParallelShape(tp=1, cp=2, ep=2), + ParallelShape(tp=1, cp=4), + ParallelShape(tp=1, cp=4, ep=2), + ParallelShape(tp=1, cp=4, ep=4), + ParallelShape(tp=2, cp=1), + ParallelShape(tp=2, cp=1, ep=2), + ), +) + +# Dense attention classes measured on the shape lattice (2026-09-04): +# Qwen3-1.7B (hidden 2,048), Qwen3-8B (4,096) and Qwen3-14B (5,120), each +# with 8 query groups of 128 channels, on H200 bf16. TP1 x CP4 is admitted +# through the two-stage re-ranker: the ten-term score cannot see the +# context-parallel plan a layout produces, which decides the recurring +# real-data groups there (design brief), so it only shortlists and the plan's +# remote wave count and largest per-rank token load select. TP2 x CP2 fails +# its gates for the two smaller classes and keeps the version-1 score. +_ATTENTION_DTYPES = ("torch.bfloat16",) +_ATTENTION_SHAPES = ( + ParallelShape(tp=1, cp=1), + ParallelShape(tp=1, cp=2), + ParallelShape(tp=2, cp=1), +) +QWEN3_1_7B_GEOMETRY = ModelGeometry( + hidden_size=2_048, + ffn_hidden_size=6_144, + num_attention_heads=16, + num_query_groups=8, + kv_channels=128, +) +DENSE_ATTN_H2048_TABLE = CalibratedTable( + table_id="dense-attn-h2048-h200-bf16", + coefficients_milli_us={ + "attention_token_cp_exchange": 0, + "gdn_level": 0, + "gdn_level_tp": 0, + "gdn_token_per_rank": 0, + "level_cp_per_layer": 0, + "level_tp_per_layer": 0, + "tiny_segment_per_layer": 0, + "token_cp_exchange": 46, + "token_per_rank": 259, + "token_tp_collective": 573, + }, + device_classes=(H200_CLASS,), + param_dtypes=_ATTENTION_DTYPES, + geometries=(QWEN3_1_7B_GEOMETRY,), + shapes=_ATTENTION_SHAPES, + reranked_shapes=(ParallelShape(tp=1, cp=4),), + reranker=ReRanker( + shortlist_size=3, + incumbent="depth_one", + shortlist_coefficients_milli_us={ + "attention_token_cp_exchange": 3, + "gdn_level": 0, + "gdn_level_tp": 0, + "gdn_token_per_rank": 0, + "level_cp_per_layer": 69_746, + "level_tp_per_layer": 0, + "tiny_segment_per_layer": 0, + "token_cp_exchange": 57, + "token_per_rank": 1_656, + "token_tp_collective": 143, + }, + wave_per_layer_milli_us=1_144_207, + max_rank_token_per_layer_milli_us=678, + ), +) +QWEN3_8B_GEOMETRY = ModelGeometry( + hidden_size=4_096, + ffn_hidden_size=12_288, + num_attention_heads=32, + num_query_groups=8, + kv_channels=128, +) +DENSE_ATTN_H4096_TABLE = CalibratedTable( + table_id="dense-attn-h4096-h200-bf16", + coefficients_milli_us={ + "attention_token_cp_exchange": 0, + "gdn_level": 0, + "gdn_level_tp": 0, + "gdn_token_per_rank": 0, + "level_cp_per_layer": 0, + "level_tp_per_layer": 0, + "tiny_segment_per_layer": 0, + "token_cp_exchange": 262, + "token_per_rank": 2_858, + "token_tp_collective": 640, + }, + device_classes=(H200_CLASS,), + param_dtypes=_ATTENTION_DTYPES, + geometries=(QWEN3_8B_GEOMETRY,), + shapes=_ATTENTION_SHAPES, + reranked_shapes=(ParallelShape(tp=1, cp=4),), + reranker=ReRanker( + shortlist_size=3, + incumbent="depth_one", + shortlist_coefficients_milli_us={ + "attention_token_cp_exchange": 57, + "gdn_level": 0, + "gdn_level_tp": 0, + "gdn_token_per_rank": 0, + "level_cp_per_layer": 17_614, + "level_tp_per_layer": 0, + "tiny_segment_per_layer": 1_508, + "token_cp_exchange": 0, + "token_per_rank": 2_925, + "token_tp_collective": 573, + }, + wave_per_layer_milli_us=249_653, + max_rank_token_per_layer_milli_us=2_820, + ), +) +QWEN3_14B_GEOMETRY = ModelGeometry( + hidden_size=5_120, + ffn_hidden_size=17_408, + num_attention_heads=40, + num_query_groups=8, + kv_channels=128, +) +DENSE_ATTN_H5120_TABLE = CalibratedTable( + table_id="dense-attn-h5120-h200-bf16", + coefficients_milli_us={ + "attention_token_cp_exchange": 0, + "gdn_level": 0, + "gdn_level_tp": 0, + "gdn_token_per_rank": 0, + "level_cp_per_layer": 0, + "level_tp_per_layer": 0, + "tiny_segment_per_layer": 0, + "token_cp_exchange": 0, + "token_per_rank": 4_607, + "token_tp_collective": 487, + }, + device_classes=(H200_CLASS,), + param_dtypes=_ATTENTION_DTYPES, + geometries=(QWEN3_14B_GEOMETRY,), + shapes=_ATTENTION_SHAPES + (ParallelShape(tp=2, cp=2),), + reranked_shapes=(ParallelShape(tp=1, cp=4),), + reranker=ReRanker( + shortlist_size=3, + incumbent="depth_one", + shortlist_coefficients_milli_us={ + "attention_token_cp_exchange": 0, + "gdn_level": 0, + "gdn_level_tp": 0, + "gdn_token_per_rank": 0, + "level_cp_per_layer": 0, + "level_tp_per_layer": 0, + "tiny_segment_per_layer": 0, + "token_cp_exchange": 0, + "token_per_rank": 4_592, + "token_tp_collective": 496, + }, + wave_per_layer_milli_us=654_967, + max_rank_token_per_layer_milli_us=4_642, + ), +) + +# Qwen3-30B-A3B class (attention + MoE, hidden 2,048): 32 attention heads in +# 4 query groups of 128 channels, 128 experts (top-8, expert FFN 768), on H200 +# bf16 (2026-09-04/05). Admitted directly at TP1 x CP1, TP1 x CP2 (EP1, EP2) +# and TP2 x CP1 (EP1, EP2); TP1 x CP4 at every expert parallelism through the +# two-stage re-ranker (this attention class has the CP4 blind spot). The CP2 +# and CP4 EP1/EP4 shapes were re-measured with sixteen rounds after their +# eight-round timings carried sporadic multi-second stalls (design brief). +ATTN_MOE_H2048_GEOMETRY = ModelGeometry( + hidden_size=2_048, + ffn_hidden_size=6_144, + num_attention_heads=32, + num_query_groups=4, + kv_channels=128, + moe_experts=128, + moe_topk=8, + moe_ffn_hidden_size=768, + moe_shared_expert_ffn=0, +) +ATTN_MOE_H2048_TABLE = CalibratedTable( + table_id="attn-moe-h2048-h200-bf16", + coefficients_milli_us={ + "attention_token_cp_exchange": 0, + "gdn_level": 0, + "gdn_level_tp": 0, + "gdn_token_per_rank": 0, + "level_cp_per_layer": 127_543, + "level_tp_per_layer": 0, + "tiny_segment_per_layer": 0, + "token_cp_exchange": 237, + "token_per_rank": 6_648, + "token_tp_collective": 105, + }, + device_classes=(H200_CLASS,), + param_dtypes=("torch.bfloat16",), + geometries=(ATTN_MOE_H2048_GEOMETRY,), + shapes=( + ParallelShape(tp=1, cp=1), + ParallelShape(tp=1, cp=2), + ParallelShape(tp=1, cp=2, ep=2), + ParallelShape(tp=2, cp=1), + ParallelShape(tp=2, cp=1, ep=2), + ), + reranked_shapes=( + ParallelShape(tp=1, cp=4), + ParallelShape(tp=1, cp=4, ep=2), + ParallelShape(tp=1, cp=4, ep=4), + ), + reranker=ReRanker( + shortlist_size=3, + incumbent="depth_one", + shortlist_coefficients_milli_us={ + "attention_token_cp_exchange": 8, + "gdn_level": 0, + "gdn_level_tp": 0, + "gdn_token_per_rank": 0, + "level_cp_per_layer": 163_376, + "level_tp_per_layer": 140_748, + "tiny_segment_per_layer": 169_817, + "token_cp_exchange": 98, + "token_per_rank": 6_705, + "token_tp_collective": 78, + }, + wave_per_layer_milli_us=1_388_921, + max_rank_token_per_layer_milli_us=1_347, + ), +) + +# Qwen3.5-27B class (dense GDN + attention, hidden 5,120): 24 attention heads +# in 4 query groups of 256 channels, GDN 48 value / 16 key heads of 128 dims +# and a 4-tap conv, 64 layers (48 GDN), on H200 bf16 (2026-09-04). Admitted at +# TP1 x CP1/2/4 and TP2 x CP1 (the ten terms rank this GDN class at CP4, as +# they do the 35B GDN MoE class); TP2 x CP2 fails its gates and keeps the +# version-1 score, which on this class lost up to 112% at CP4. +QWEN35_27B_GEOMETRY = ModelGeometry( + hidden_size=5_120, + ffn_hidden_size=17_408, + num_attention_heads=24, + num_query_groups=4, + kv_channels=256, + gdn_value_heads=48, + gdn_key_heads=16, + gdn_key_head_dim=128, + gdn_value_head_dim=128, + gdn_conv_kernel=4, +) +DENSE_GDN_H5120_TABLE = CalibratedTable( + table_id="dense-gdn-h5120-h200-bf16", + coefficients_milli_us={ + "attention_token_cp_exchange": 0, + "gdn_level": 3_043_381, + "gdn_level_tp": 0, + "gdn_token_per_rank": 0, + "level_cp_per_layer": 0, + "level_tp_per_layer": 442_039, + "tiny_segment_per_layer": 230_182, + "token_cp_exchange": 0, + "token_per_rank": 11_441, + "token_tp_collective": 0, + }, + device_classes=(H200_CLASS,), + param_dtypes=("torch.bfloat16",), + geometries=(QWEN35_27B_GEOMETRY,), + shapes=( + ParallelShape(tp=1, cp=1), + ParallelShape(tp=1, cp=2), + ParallelShape(tp=1, cp=4), + ParallelShape(tp=2, cp=1), + ), +) + +CALIBRATED_TABLES: tuple[CalibratedTable, ...] = ( + DENSE_H2560_TABLE, + GDN_MOE_H2048_TABLE, + DENSE_ATTN_H2048_TABLE, + DENSE_ATTN_H4096_TABLE, + DENSE_ATTN_H5120_TABLE, + ATTN_MOE_H2048_TABLE, + DENSE_GDN_H5120_TABLE, ) -CALIBRATED_TABLES: tuple[CalibratedTable, ...] = (DENSE_H2560_TABLE,) # CPU-only planning (unit tests) prices with this table. DEFAULT_TABLE = DENSE_H2560_TABLE assert set(COEFFICIENTS_MILLI_US) == set(TERM_FUNCTIONS) +assert all( + set(table.coefficients_milli_us) == set(TERM_FUNCTIONS) + for table in CALIBRATED_TABLES +) def predicted_work( @@ -580,12 +985,19 @@ def prefix_tree_layout_score( "COEFFICIENT_VERSION", "COEFFICIENT_VERSION_FALLBACK", "DEFAULT_TABLE", + "ATTN_MOE_H2048_TABLE", + "DENSE_ATTN_H2048_TABLE", + "DENSE_ATTN_H4096_TABLE", + "DENSE_ATTN_H5120_TABLE", + "DENSE_GDN_H5120_TABLE", "DENSE_H2560_TABLE", + "GDN_MOE_H2048_TABLE", "CalibratedTable", "DeviceClass", "LayoutFeatures", "ModelGeometry", "ParallelShape", + "ReRanker", "ScoringSelection", "SEGMENT_LENGTH_THRESHOLDS", "ScoringFacts", diff --git a/src/art/trainer_rank/_prefix_tree_planner.py b/src/art/trainer_rank/_prefix_tree_planner.py index 7abf285cc..22f604eb5 100644 --- a/src/art/trainer_rank/_prefix_tree_planner.py +++ b/src/art/trainer_rank/_prefix_tree_planner.py @@ -25,10 +25,13 @@ from itertools import combinations import json import struct -from typing import Literal, TypeAlias, overload +from typing import TYPE_CHECKING, Literal, TypeAlias, overload import torch +if TYPE_CHECKING: + from ._planner_cost import ReRanker + Fingerprint: TypeAlias = str DecisionSet: TypeAlias = frozenset[int] FixedPointScore: TypeAlias = int | tuple[int, ...] @@ -755,6 +758,8 @@ def select_prefix_tree_layout( gdn_layers: int | None = None, coefficient_version: int | None = None, coefficients: Mapping[str, int] | None = None, + reranker: "ReRanker | None" = None, + plan_structure: Callable[[PrefixTreeLayout], tuple[int, int]] | None = None, ): """Production layout selection for one canonical tree. @@ -764,9 +769,20 @@ def select_prefix_tree_layout( ``gdn_layers`` defaults to ``layers`` when the model uses GDN; ``coefficients`` is the calibrated table for this runtime's execution class (default: the shipped dense table). + + With a ``reranker`` (execution classes whose shape is re-ranked, see + ``select_scoring``) the cheap score only shortlists: its + ``shortlist_size`` best layouts plus the ``incumbent`` anchor are priced + by the context-parallel plan each produces, ``plan_structure(layout)`` = + (remote wave count, largest per-rank token load), and the lowest + second-stage score wins; ties keep the cheaper layout. """ from ._planner_cost import COEFFICIENT_VERSION, prefix_tree_layout_score + from ._prefix_tree_performance_search import ( + _candidate_rank_key, + search_nonuniform_prefix_tree_layouts, + ) version = ( COEFFICIENT_VERSION if coefficient_version is None else coefficient_version @@ -784,12 +800,38 @@ def scorer(layout: PrefixTreeLayout) -> tuple[int, ...]: coefficients=coefficients, ) - return search_prefix_tree_layout( + result = search_nonuniform_prefix_tree_layouts( tree, scorer, refinement_work_budget=refinement_work_budget, mandatory_candidates=prefix_tree_layout_candidates(tree), ) + if reranker is None: + return result.incumbent + if plan_structure is None: + raise ValueError("re-ranking needs the plan structure of a layout") + ranked = sorted(result.candidates, key=_candidate_rank_key) + shortlist = list(ranked[: max(1, reranker.shortlist_size)]) + shortlist += [ + anchor + for anchor in ranked + if anchor.mandatory + and reranker.incumbent in anchor.candidate.labels + and anchor not in shortlist + ] + best: tuple[tuple[int, int], object] | None = None + for position, candidate in enumerate(shortlist): + wave_count, max_rank_tokens = plan_structure(candidate.layout) + key = ( + reranker.score( + layers=layers, wave_count=wave_count, max_rank_tokens=max_rank_tokens + ), + position, + ) + if best is None or key < best[0]: + best = (key, candidate) + assert best is not None + return best[1] __all__ = [ diff --git a/tests/unit/test_context_parallel_plan_evaluation.py b/tests/unit/test_context_parallel_plan_evaluation.py new file mode 100644 index 000000000..db0057e24 --- /dev/null +++ b/tests/unit/test_context_parallel_plan_evaluation.py @@ -0,0 +1,446 @@ +"""The vectorized CP plan evaluator reproduces the scalar reference exactly. + +``_evaluate_plan`` prices one chunk assignment (per-rank forward and backward +time) inside the context-parallel assignment search. It was rewritten from +per-rank torch mask operations into numpy mask algebra over all ranks, sources +and waves at once (about six times cheaper per evaluation); the scalar cost +formulas and stage simulations are shared. The pre-rewrite implementation is +kept here as the reference: every randomized instance below must price +identically, floats included. +""" + +from __future__ import annotations + +import random +from typing import Any + +import pytest +import torch + +pytest.importorskip("megatron.core") +from art.megatron.context_parallel import runtime as rt # noqa: E402 +from art.megatron.context_parallel.runtime import ( # noqa: E402 + _comm_cost_ms, + _simulate_backward_time_ms, + _simulate_forward_time_ms, + _stage_cost_ms, +) +from art.megatron.context_parallel.types import ( # noqa: E402 + ContextParallelConfig, + TokenRange, +) + +_CHUNK_MASK_STATS_TORCH_THRESHOLD = 1024 + + +def _reference_chunk_mask_stats( + *, + chunk_lengths: tuple[int, ...], + chunk_mask: torch.Tensor, + chunk_lengths_tensor: torch.Tensor | None = None, +) -> tuple[int, int]: + if ( + chunk_lengths_tensor is not None + and len(chunk_lengths) >= _CHUNK_MASK_STATS_TORCH_THRESHOLD + ): + if int(chunk_mask.numel()) == 0 or not bool(chunk_mask.any().item()): + return 0, 0 + token_count = int(chunk_lengths_tensor[chunk_mask].sum().item()) + run_starts = chunk_mask.clone() + run_starts[1:] = torch.logical_and( + run_starts[1:], torch.logical_not(chunk_mask[:-1]) + ) + range_count = int(run_starts.sum().item()) + return token_count, range_count + token_count = 0 + range_count = 0 + in_run = False + for is_set, length in zip(chunk_mask.tolist(), chunk_lengths, strict=True): + if bool(is_set): + token_count += int(length) + if not in_run: + range_count += 1 + in_run = True + continue + in_run = False + return token_count, range_count + + +def _reference_evaluate_plan( + *, + chunk_ranges: tuple[TokenRange, ...], + pair_matrix: list[list[int]] | torch.Tensor, + owners: tuple[int, ...], + wave_assignment: tuple[int, ...], + cp_size: int, + config: ContextParallelConfig, + pair_positive: torch.Tensor | None = None, + chunk_lengths: tuple[int, ...] | None = None, + chunk_lengths_tensor: torch.Tensor | None = None, +) -> dict[str, Any]: + rank_scores: list[float] = [] + rank_forward_ms: list[float] = [] + rank_backward_ms: list[float] = [] + chunk_count = len(chunk_ranges) + wave_count = max(wave_assignment, default=0) + 1 if wave_assignment else 0 + pair_counts = ( + pair_matrix + if isinstance(pair_matrix, torch.Tensor) and pair_matrix.dtype == torch.int64 + else torch.as_tensor(pair_matrix, dtype=torch.int64) + ) + if pair_positive is None: + pair_positive = pair_counts > 0 + if chunk_lengths is None: + chunk_lengths = tuple(int(range_.size()) for range_ in chunk_ranges) + if ( + chunk_lengths_tensor is None + and len(chunk_lengths) >= _CHUNK_MASK_STATS_TORCH_THRESHOLD + ): + chunk_lengths_tensor = torch.tensor(chunk_lengths, dtype=torch.int64) + owners_tensor = torch.tensor(owners, dtype=torch.int64) + wave_tensor = torch.tensor( + wave_assignment, + dtype=torch.int64, + ) + owner_masks = [owners_tensor == rank for rank in range(cp_size)] + owner_indices = [ + torch.nonzero(owner_mask, as_tuple=False).flatten() + for owner_mask in owner_masks + ] + empty_pair_counts = pair_counts.new_zeros((0, chunk_count)) + empty_pair_positive = pair_positive.new_zeros((0, chunk_count)) + pair_counts_by_rank_rows = [ + ( + empty_pair_counts + if int(owner_index.numel()) == 0 + else pair_counts.index_select(0, owner_index) + ) + for owner_index in owner_indices + ] + pair_positive_by_rank_rows = [ + ( + empty_pair_positive + if int(owner_index.numel()) == 0 + else pair_positive.index_select(0, owner_index) + ) + for owner_index in owner_indices + ] + pair_positive_by_rank_cols = [ + ( + torch.zeros(chunk_count, dtype=torch.bool) + if int(rank_rows.numel()) == 0 + else rank_rows.any(dim=0) + ) + for rank_rows in pair_positive_by_rank_rows + ] + wave_masks = [wave_tensor == wave_index for wave_index in range(wave_count)] + + for rank in range(cp_size): + owned_q_mask = owner_masks[rank] + owned_q_indices = owner_indices[rank] + owned_pair_counts = pair_counts_by_rank_rows[rank] + owned_pair_positive = pair_positive_by_rank_rows[rank] + owned_positive_cols = pair_positive_by_rank_cols[rank] + + local_pairs = ( + 0 + if int(owned_q_indices.numel()) == 0 + else int(owned_pair_counts.index_select(1, owned_q_indices).sum().item()) + ) + local_q_mask = torch.zeros(chunk_count, dtype=torch.bool) + if int(owned_q_indices.numel()) > 0: + touched_local_q = owned_pair_positive.index_select(1, owned_q_indices).any( + dim=1 + ) + if bool(touched_local_q.any().item()): + local_q_mask[owned_q_indices[touched_local_q]] = True + local_k_mask = owned_q_mask & owned_positive_cols + local_q_tokens, local_q_range_count = _reference_chunk_mask_stats( + chunk_lengths=chunk_lengths, + chunk_mask=local_q_mask, + chunk_lengths_tensor=chunk_lengths_tensor, + ) + local_k_tokens, local_k_range_count = _reference_chunk_mask_stats( + chunk_lengths=chunk_lengths, + chunk_mask=local_k_mask, + chunk_lengths_tensor=chunk_lengths_tensor, + ) + local_stage_ms = _stage_cost_ms( + pair_count=local_pairs, + q_tokens=local_q_tokens, + k_tokens=local_k_tokens, + q_range_count=local_q_range_count, + k_range_count=local_k_range_count, + config=config, + backward=False, + local=True, + ) + local_backward_ms = _stage_cost_ms( + pair_count=local_pairs, + q_tokens=local_q_tokens, + k_tokens=local_k_tokens, + q_range_count=local_q_range_count, + k_range_count=local_k_range_count, + config=config, + backward=True, + local=True, + ) + + remote_stage_ms: list[float] = [] + remote_fetch_ms: list[float] = [] + remote_backward_ms: list[float] = [] + remote_reduce_ms: list[float] = [] + for wave_index in range(wave_count): + request_tokens_by_source = [0 for _ in range(cp_size)] + request_range_counts_by_source = [0 for _ in range(cp_size)] + request_pairs = 0 + touched_q_mask = torch.zeros(chunk_count, dtype=torch.bool) + for source_rank in range(cp_size): + if source_rank == rank: + continue + touched_source_mask = ( + owner_masks[source_rank] + & wave_masks[wave_index] + & owned_positive_cols + ) + ( + request_tokens_by_source[source_rank], + request_range_counts_by_source[source_rank], + ) = _reference_chunk_mask_stats( + chunk_lengths=chunk_lengths, + chunk_mask=touched_source_mask, + chunk_lengths_tensor=chunk_lengths_tensor, + ) + if request_tokens_by_source[source_rank] <= 0: + continue + touched_source_indices = torch.nonzero( + touched_source_mask, + as_tuple=False, + ).flatten() + request_pairs += int( + owned_pair_counts.index_select(1, touched_source_indices) + .sum() + .item() + ) + touched_remote_q = owned_pair_positive.index_select( + 1, + touched_source_indices, + ).any(dim=1) + if bool(touched_remote_q.any().item()): + touched_q_mask[owned_q_indices[touched_remote_q]] = True + recv_tokens = sum(request_tokens_by_source) + recv_range_count = sum(request_range_counts_by_source) + if request_pairs <= 0 and recv_tokens <= 0 and recv_range_count <= 0: + continue + + send_tokens_by_peer = [0 for _ in range(cp_size)] + send_range_counts_by_peer = [0 for _ in range(cp_size)] + aggregate_send_mask = torch.zeros(chunk_count, dtype=torch.bool) + owned_wave_mask = owned_q_mask & wave_masks[wave_index] + if bool(owned_wave_mask.any().item()): + for peer_rank in range(cp_size): + if peer_rank == rank: + continue + send_mask = owned_wave_mask & pair_positive_by_rank_cols[peer_rank] + ( + send_tokens_by_peer[peer_rank], + send_range_counts_by_peer[peer_rank], + ) = _reference_chunk_mask_stats( + chunk_lengths=chunk_lengths, + chunk_mask=send_mask, + chunk_lengths_tensor=chunk_lengths_tensor, + ) + if send_tokens_by_peer[peer_rank] > 0: + aggregate_send_mask |= send_mask + ( + send_tokens_by_peer[rank], + send_range_counts_by_peer[rank], + ) = _reference_chunk_mask_stats( + chunk_lengths=chunk_lengths, + chunk_mask=aggregate_send_mask, + chunk_lengths_tensor=chunk_lengths_tensor, + ) + + send_tokens = sum(send_tokens_by_peer) + q_tokens, q_range_count = _reference_chunk_mask_stats( + chunk_lengths=chunk_lengths, + chunk_mask=touched_q_mask, + chunk_lengths_tensor=chunk_lengths_tensor, + ) + remote_stage_ms.append( + _stage_cost_ms( + pair_count=request_pairs, + q_tokens=q_tokens, + k_tokens=recv_tokens, + q_range_count=q_range_count, + k_range_count=recv_range_count, + config=config, + backward=False, + local=False, + ) + ) + remote_backward_ms.append( + _stage_cost_ms( + pair_count=request_pairs, + q_tokens=q_tokens, + k_tokens=recv_tokens, + q_range_count=q_range_count, + k_range_count=recv_range_count, + config=config, + backward=True, + local=False, + ) + ) + remote_fetch_ms.append( + _comm_cost_ms( + tokens=max(send_tokens, recv_tokens), + range_count=max(sum(send_range_counts_by_peer), recv_range_count), + config=config, + backward=False, + ) + ) + remote_reduce_ms.append( + _comm_cost_ms( + tokens=max(send_tokens, recv_tokens), + range_count=max(sum(send_range_counts_by_peer), recv_range_count), + config=config, + backward=True, + ) + ) + + forward_ms = _simulate_forward_time_ms( + local_stage_ms=local_stage_ms if local_pairs > 0 else 0.0, + remote_stage_ms=tuple(remote_stage_ms), + remote_fetch_ms=tuple(remote_fetch_ms), + ) + backward_ms = _simulate_backward_time_ms( + local_stage_ms=local_backward_ms if local_pairs > 0 else 0.0, + remote_stage_ms=tuple(remote_backward_ms), + remote_reduce_ms=tuple(remote_reduce_ms), + ) + rank_forward_ms.append(float(forward_ms)) + rank_backward_ms.append(float(backward_ms)) + rank_scores.append(float(forward_ms + backward_ms)) + return { + "score": max(rank_scores, default=0.0), + "rank_scores": tuple(rank_scores), + "rank_forward_ms": tuple(rank_forward_ms), + "rank_backward_ms": tuple(rank_backward_ms), + } + + +def _random_instance(seed: int, *, chunk_count: int | None = None): + rng = random.Random(seed) + cp_size = rng.choice((1, 2, 4, 8)) + n = rng.randint(0, 48) if chunk_count is None else chunk_count + chunk_size = 512 + lengths = [chunk_size] * n + if n: + lengths[-1] = rng.randint(1, chunk_size) + ranges, start = [], 0 + for length in lengths: + ranges.append(TokenRange(start=start, end=start + length)) + start += length + pairs = [[0] * n for _ in range(n)] + density = rng.choice((0.15, 0.4, 0.8)) + for i in range(n): + if rng.random() < 0.1: + continue # a q chunk that attends nothing (e.g. padding) + for j in range(i + 1): + if rng.random() < density or j == i: + pairs[i][j] = rng.randint(1, chunk_size * chunk_size) + owners = tuple(rng.randrange(cp_size) for _ in range(n)) + wave_count = rng.randint(1, 4) + waves = rt._wave_assignment(chunk_count=n, wave_count=wave_count) + return tuple(ranges), pairs, owners, waves, cp_size + + +def _both( + ranges, pairs, owners, waves, cp_size +) -> tuple[dict[str, Any], dict[str, Any]]: + config = ContextParallelConfig() + program = rt._pair_program(pairs, chunk_ranges=ranges) + new = rt._evaluate_plan( + program=program, + owners=owners, + wave_assignment=waves, + cp_size=cp_size, + config=config, + ) + reference = _reference_evaluate_plan( + chunk_ranges=ranges, + pair_matrix=torch.as_tensor(pairs, dtype=torch.int64).reshape( + len(ranges), len(ranges) + ), + owners=owners, + wave_assignment=waves, + cp_size=cp_size, + config=config, + ) + return new, reference + + +@pytest.mark.parametrize("seed", range(120)) +def test_vectorized_evaluator_matches_scalar_reference(seed: int) -> None: + new, reference = _both(*_random_instance(seed)) + assert new == reference + + +def test_vectorized_evaluator_matches_reference_on_a_long_row() -> None: + # The reference switches to its torch path beyond 1,024 chunks. + new, reference = _both(*_random_instance(7, chunk_count=1_100)) + assert new == reference + + +@pytest.mark.parametrize("seed", range(200, 230)) +def test_batched_evaluation_matches_one_by_one(seed: int) -> None: + import numpy as np + + ranges, pairs, _owners, waves, cp_size = _random_instance(seed) + rng = random.Random(seed) + batch = [ + tuple(rng.randrange(cp_size) for _ in range(len(ranges))) for _ in range(7) + ] + program = rt._pair_program(pairs, chunk_ranges=ranges) + config = ContextParallelConfig() + together = rt._evaluate_plans( + program=program, + owners_batch=np.asarray(batch, dtype=np.int64).reshape(len(batch), len(ranges)), + wave_assignment=waves, + cp_size=cp_size, + config=config, + ) + for owners, evaluation in zip(batch, together, strict=True): + assert evaluation == _reference_evaluate_plan( + chunk_ranges=ranges, + pair_matrix=torch.as_tensor(pairs, dtype=torch.int64).reshape( + len(ranges), len(ranges) + ), + owners=owners, + wave_assignment=waves, + cp_size=cp_size, + config=config, + ) + + +def test_empty_row_prices_to_zero() -> None: + new, reference = _both((), [], (), (), 4) + assert new == reference + assert new["score"] == 0.0 and len(new["rank_scores"]) == 4 + + +def test_search_is_unchanged_by_the_rewrite() -> None: + """The assignment search only consumes evaluations; on a fixed program its + result (owners, waves, score) is a pure function of them.""" + + ranges, pairs, _owners, _waves, cp_size = _random_instance(11, chunk_count=40) + q_weights = [float(sum(row)) for row in pairs] + owners, waves, evaluation = rt._search_generic_chunk_assignment( + chunk_ranges=ranges, + pair_matrix=torch.as_tensor(pairs, dtype=torch.int64), + q_weights=q_weights, + cp_size=max(cp_size, 2), + config=ContextParallelConfig(), + ) + _new, reference = _both(ranges, pairs, owners, waves, max(cp_size, 2)) + assert evaluation == reference diff --git a/tests/unit/test_planner_cost_certificate.py b/tests/unit/test_planner_cost_certificate.py index 74c709715..71940ebab 100644 --- a/tests/unit/test_planner_cost_certificate.py +++ b/tests/unit/test_planner_cost_certificate.py @@ -1,12 +1,15 @@ -"""The checked-in calibration certificate binds the production table to its data. - -The certificate carries every cell's candidate features, median timings, -counts and fingerprints (no tokens), the exact fit arguments, the integer table -and its hash. This test verifies (cheaply) that the shipped table is the -certificate's table and that the certificate's headline metrics are what the -table produces on the recorded aggregates. Set ``ART_COST_CERTIFICATE_REFIT=1`` -to also re-run the full fit from the aggregates and require the identical -integer table (slower: about a minute). +"""Each checked-in calibration certificate binds one production table to its data. + +One certificate per calibrated table +(``dev/trainer_rank_cost_calibration_certificate_
.json``). A certificate +carries every cell's candidate features, median timings, counts and execution +fingerprints (no tokens), the exact fit arguments, the integer table and its +hash. These tests verify (cheaply) that each shipped table is its certificate's +table, that the certified headline metrics are what the table produces on the +recorded aggregates, and that the table admits exactly the execution classes +the certificate measured. Set ``ART_COST_CERTIFICATE_REFIT=1`` to also re-run +the full fit from the aggregates and require the identical integer table +(slower: about a minute per table). """ from __future__ import annotations @@ -19,6 +22,7 @@ import sys import numpy as np +import pytest from art.trainer_rank._planner_cost import ( CALIBRATED_TABLES, @@ -30,8 +34,6 @@ ) _ROOT = Path(__file__).resolve().parents[2] -_CERTIFICATE = _ROOT / "dev" / "trainer_rank_cost_calibration_certificate.json" -_MANIFEST = _ROOT / "dev" / "trainer_rank_cost_calibration_manifest.json" _spec = importlib.util.spec_from_file_location( "trainer_rank_cost_fit", _ROOT / "dev" / "trainer_rank_cost_fit.py" ) @@ -41,26 +43,63 @@ _spec.loader.exec_module(fit) -def test_shipped_table_is_the_certified_table() -> None: - payload = json.loads(_CERTIFICATE.read_text()) - assert payload["schema"] == fit.CERTIFICATE_SCHEMA - assert payload["table_id"] == DENSE_H2560_TABLE.table_id - assert payload["integer_table_milli_us"] == COEFFICIENTS_MILLI_US +def certificate_path(table_id: str) -> Path: + return _ROOT / "dev" / f"trainer_rank_cost_calibration_certificate_{table_id}.json" + + +def manifest_path(table_id: str) -> Path: + return _ROOT / "dev" / f"trainer_rank_cost_calibration_manifest_{table_id}.json" + + +_TABLES = pytest.mark.parametrize("table", CALIBRATED_TABLES, ids=lambda t: t.table_id) + + +def _label(shape: ParallelShape) -> str: + """The fitter's shape label (``_shape_label``): tp, cp, then ep/etp when > 1.""" + + return ( + f"tp{shape.tp}cp{shape.cp}" + + (f"ep{shape.ep}" if shape.ep > 1 else "") + + (f"etp{shape.etp}" if shape.etp > 1 else "") + ) + + +def test_default_table_is_the_dense_one() -> None: assert DENSE_H2560_TABLE.coefficients_milli_us == COEFFICIENTS_MILLI_US - assert [table.table_id for table in CALIBRATED_TABLES] == [payload["table_id"]] + assert len({table.table_id for table in CALIBRATED_TABLES}) == len( + CALIBRATED_TABLES + ) + + +@_TABLES +def test_shipped_table_is_the_certified_table(table) -> None: + payload = json.loads(certificate_path(table.table_id).read_text()) + assert payload["schema"] == fit.CERTIFICATE_SCHEMA + assert payload["table_id"] == table.table_id + assert payload["integer_table_milli_us"] == dict(table.coefficients_milli_us) digest = hashlib.sha256( json.dumps(payload["integer_table_milli_us"], sort_keys=True).encode() ).hexdigest() assert digest == payload["integer_table_sha256"] -def test_certified_metrics_hold_on_the_recorded_aggregates() -> None: - candidates, payload = fit.load_certificate(_CERTIFICATE) +@_TABLES +def test_certified_metrics_hold_on_the_recorded_aggregates(table) -> None: + candidates, payload = fit.load_certificate(certificate_path(table.table_id)) terms = tuple(payload["fit_arguments"]["terms"]) - beta = np.asarray([COEFFICIENTS_MILLI_US[name] / 1_000.0 for name in terms]) + beta = np.asarray( + [table.coefficients_milli_us.get(name, 0) / 1_000.0 for name in terms] + ) + reranked_labels = {_label(shape) for shape in table.reranked_shapes} + candidates = [c for c in candidates if fit._shape_label(c) not in reranked_labels] report = fit.evaluate(candidates, fit.predict(candidates, terms, beta)) recorded = payload["metrics"]["integer_all"] - assert report["cells"] == recorded["cells"] == len(payload["cells"]) + direct_cells = [ + cell + for cell in payload["cells"] + if ParallelShape(*cell["shape"]) not in table.reranked_shapes + ] + assert report["cells"] == recorded["cells"] == len(direct_cells) assert report["ordered_pairs"] == recorded["ordered_pairs"] assert abs(report["pairwise_accuracy"] - recorded["pairwise_accuracy"]) < 1e-9 assert abs(report["max_regret_pct"] - recorded["max_regret_pct"]) < 1e-9 @@ -69,10 +108,11 @@ def test_certified_metrics_hold_on_the_recorded_aggregates() -> None: assert not report["clear_misses"] -def test_full_refit_reproduces_the_table_when_requested() -> None: +@_TABLES +def test_full_refit_reproduces_the_table_when_requested(table) -> None: if os.environ.get("ART_COST_CERTIFICATE_REFIT") != "1": return - candidates, payload = fit.load_certificate(_CERTIFICATE) + candidates, payload = fit.load_certificate(certificate_path(table.table_id)) arguments = payload["fit_arguments"] terms = tuple(arguments["terms"]) @@ -86,30 +126,103 @@ def held_out(cell: str) -> bool: and int(cell.rsplit("|g", 1)[1]) % 2 == 1 ) - train = [c for c in candidates if not held_out(c.cell)] + reranked_labels = {_label(shape) for shape in table.reranked_shapes} + train = [ + c + for c in candidates + if not held_out(c.cell) and fit._shape_label(c) not in reranked_labels + ] beta = fit.nnls(*fit.paired_deltas(train, terms)) if arguments["objective"] == "regret": beta = fit.fit_regret(train, terms, beta) - table = { - name: int(round(value * 1_000)) for name, value in zip(terms, beta, strict=True) + assert fit.integerize(train, terms, beta) == payload["integer_table_milli_us"] + + +@_TABLES +def test_reranked_shapes_are_certified_by_the_two_stage_selection(table) -> None: + """Tables with a re-ranker: the shipped re-ranker is the certificate's, and + the two-stage selection recomputed from the recorded aggregates (shortlist + by the shortlist table, select by the plan structure) reproduces the + certified metrics and passes the gates on every re-ranked shape; the + direct metrics cover the directly scored shapes only.""" + + candidates, payload = fit.load_certificate(certificate_path(table.table_id)) + block = payload.get("reranker") + if table.reranker is None: + assert block is None and not table.reranked_shapes + assert all( + ParallelShape(*cell["shape"]) in table.shapes for cell in payload["cells"] + ) + return + reranked_labels = {_label(shape) for shape in table.reranked_shapes} + assert set(block["shapes"]) == reranked_labels + assert block["shortlist_size"] == table.reranker.shortlist_size + assert block["incumbent"] == table.reranker.incumbent + assert block["shortlist_table_milli_us"] == dict( + table.reranker.shortlist_coefficients_milli_us + ) + assert block["wave_per_layer_milli_us"] == table.reranker.wave_per_layer_milli_us + assert ( + block["max_rank_token_per_layer_milli_us"] + == table.reranker.max_rank_token_per_layer_milli_us + ) + terms = tuple(payload["fit_arguments"]["terms"]) + reranked = [c for c in candidates if fit._shape_label(c) in reranked_labels] + direct = [c for c in candidates if fit._shape_label(c) not in reranked_labels] + assert reranked and direct + assert {fit._shape_label(c) for c in direct} == { + _label(shape) for shape in table.shapes } - assert table == payload["integer_table_milli_us"] + recomputed = fit.evaluate_two_stage( + reranked, + terms=terms, + shortlist_beta=np.asarray( + [block["shortlist_table_milli_us"][name] / 1_000.0 for name in terms] + ), + rerank_beta=np.asarray( + [ + block["wave_per_layer_milli_us"] / 1_000.0, + block["max_rank_token_per_layer_milli_us"] / 1_000.0, + ] + ), + shortlist_size=block["shortlist_size"], + incumbent=block["incumbent"], + ) + recorded = block["metrics"]["all"] + for key in ("cells", "ordered_pairs", "clear_misses", "recall"): + assert recomputed[key] == recorded[key], key + for key in ( + "pairwise_accuracy", + "median_regret_pct", + "p95_regret_pct", + "max_regret_pct", + ): + assert abs(recomputed[key] - recorded[key]) < 1e-9, key + assert not fit.two_stage_gates(recomputed) + for shape, metrics in block["metrics"]["by_shape"].items(): + assert not fit.two_stage_gates(metrics), shape + # The direct metrics are the certificate's headline metrics: direct cells only. + beta = np.asarray( + [table.coefficients_milli_us.get(name, 0) / 1_000.0 for name in terms] + ) + report = fit.evaluate(direct, fit.predict(direct, terms, beta)) + assert report["cells"] == payload["metrics"]["integer_all"]["cells"] -def test_certificate_holds_exactly_the_manifest_cells() -> None: - """Exact cell identities: every expected cell minus the explicit exclusions - (none since the two Ellavox CP4 cells were re-measured after issue #840).""" +@_TABLES +def test_certificate_holds_exactly_the_manifest_cells(table) -> None: + """Exact cell identities: every expected cell minus the explicit exclusions.""" - payload = json.loads(_CERTIFICATE.read_text()) - manifest = json.loads(_MANIFEST.read_text()) + payload = json.loads(certificate_path(table.table_id).read_text()) + manifest = json.loads(manifest_path(table.table_id).read_text()) + assert manifest["table_id"] == table.table_id expected = {cell["key"] for cell in manifest["cells"]} excluded = {cell["key"] for cell in manifest["excluded"]} assert excluded <= expected certified = {cell["cell"] for cell in payload["cells"]} assert certified == expected - excluded - assert len(certified) == 58 and not excluded assert set(payload["manifest"]["excluded"]) == excluded - assert payload["manifest"]["path"] == _MANIFEST.name + assert payload["manifest"]["path"] == manifest_path(table.table_id).name # Every retained cell carries its execution fingerprints. for cell in payload["cells"]: for key in ( @@ -118,46 +231,64 @@ def test_certificate_holds_exactly_the_manifest_cells() -> None: "device", "param_dtype", "hidden_size", + "geometry", + "shape", + "model", ): - assert cell.get(key) not in (None, ""), (cell["cell"], key) + assert cell.get(key) not in (None, "", [], {}), (cell["cell"], key) -def test_admitted_execution_classes_are_exactly_the_certified_ones() -> None: +@_TABLES +def test_admitted_execution_classes_are_exactly_the_certified_ones(table) -> None: """The production table admits precisely the device classes, dtypes, model geometries and parallel shapes recorded in the certificate's cells.""" - payload = json.loads(_CERTIFICATE.read_text()) + payload = json.loads(certificate_path(table.table_id).read_text()) cells = payload["cells"] + memory_classes = {device.memory_class for device in table.device_classes} for cell in cells: - assert cell["geometry"] and cell["shape"] and cell["model"], cell["cell"] - assert ( - cell["device_memory_class"] - == DeviceClass.memory_class_for( - None if cell.get("device_total_memory_bytes") is None else None - ) - or cell["device_memory_class"] == "hbm-141g" - ) + assert cell["device_memory_class"] in memory_classes, cell["cell"] admitted = payload["admitted"] geometries = {ModelGeometry(**geometry) for geometry in admitted["geometries"]} - assert geometries == set(DENSE_H2560_TABLE.geometries) + assert geometries == set(table.geometries) assert geometries == {ModelGeometry(**cell["geometry"]) for cell in cells} shapes = {ParallelShape(*shape) for shape in admitted["shapes"]} - assert shapes == set(DENSE_H2560_TABLE.shapes) + assert shapes == set(table.shapes) | set(table.reranked_shapes) assert shapes == {ParallelShape(*cell["shape"]) for cell in cells} devices = { DeviceClass(tuple(entry["capability"]), entry["memory_class"]) for entry in admitted["device_classes"] } - assert devices == set(DENSE_H2560_TABLE.device_classes) - assert set(admitted["param_dtypes"]) == set(DENSE_H2560_TABLE.param_dtypes) - # Every geometry was measured at every admitted shape (product admission). + assert devices == set(table.device_classes) + assert set(admitted["param_dtypes"]) == set(table.param_dtypes) + # Every geometry was measured at every admitted shape (product admission); + # a withheld pair is a measured pair the manifest documents with a reason, + # and the table does not admit it. measured = { (ModelGeometry(**cell["geometry"]), ParallelShape(*cell["shape"])) for cell in cells } - for geometry in DENSE_H2560_TABLE.geometries: - for shape in DENSE_H2560_TABLE.shapes: + for geometry in table.geometries: + for shape in (*table.shapes, *table.reranked_shapes): assert (geometry, shape) in measured, (geometry.hidden_size, shape) + manifest = json.loads(manifest_path(table.table_id).read_text()) + geometry_of_model = { + cell["model"]: ModelGeometry(**cell["geometry"]) for cell in cells + } + withheld = { + (geometry_of_model[entry["model"]], ParallelShape(*entry["shape"])) + for entry in manifest.get("withheld", []) + } + assert withheld == set(table.withheld) + assert withheld <= measured + assert all(entry.get("reason") for entry in manifest.get("withheld", [])) + for geometry, shape in withheld: + assert not table.admits( + device=table.device_classes[0], + param_dtype=table.param_dtypes[0], + geometry=geometry, + shape=shape, + ) # One geometry per model across all of its cells. by_model: dict[str, set[str]] = {} for cell in cells: diff --git a/tests/unit/test_planner_cost_manifest_recipes.py b/tests/unit/test_planner_cost_manifest_recipes.py index 6d5f70962..aa3e6515c 100644 --- a/tests/unit/test_planner_cost_manifest_recipes.py +++ b/tests/unit/test_planner_cost_manifest_recipes.py @@ -1,9 +1,11 @@ -"""The calibration manifest is exactly what the checked-in recipes launch. - -Parses the three runners (the local script and the two SkyPilot recipes in -both ``CELL_SET`` modes), expands the Ellavox group loops, resolves ``layers=0`` -to each model's depth, and requires the resulting cell keys to equal the -manifest's, so a clean rerun of every documented recipe reproduces the +"""Each calibration manifest is exactly what its checked-in launches produce. + +One manifest per calibrated table (``dev/trainer_rank_cost_calibration_manifest_
.json``). +A manifest names the fixed recipes it was run from (the local script and the +two SkyPilot recipes in both ``CELL_SET`` modes, parsed here with their +Ellavox group loops expanded) and/or the shape-lattice launches (model, shapes, +cells) of the generic lattice recipe; the resulting cell keys must equal the +manifest's, so a clean rerun of the documented launches reproduces the certified cell set (minus the explicit exclusions). """ @@ -13,17 +15,43 @@ from pathlib import Path import re +import pytest + +from art.trainer_rank._planner_cost import CALIBRATED_TABLES + _ROOT = Path(__file__).resolve().parents[2] _DEV = _ROOT / "dev" -_MANIFEST = _DEV / "trainer_rank_cost_calibration_manifest.json" -_FULL_LAYERS = {"Qwen/Qwen3.5-4B": 32, "Qwen/Qwen3-4B": 36} +_FULL_LAYERS = { + "Qwen/Qwen3.5-4B": 32, + "Qwen/Qwen3-4B": 36, + "Qwen/Qwen3.5-35B-A3B": 40, + "Qwen/Qwen3-8B": 36, + "Qwen/Qwen3-1.7B": 28, + "Qwen/Qwen3-14B": 40, + "Qwen/Qwen3.5-27B": 64, + "Qwen/Qwen3-30B-A3B": 48, +} _ELLAVOX_GROUPS = range(8) +_LATTICE_RECIPE = "dev/trainer_rank_cost_calibration_lattice.sky.yaml" + + +def manifest_path(table_id: str) -> Path: + return _DEV / f"trainer_rank_cost_calibration_manifest_{table_id}.json" def _key( - cell: str, model: str, layers: int, tp: int, cp: int, group: int | None + cell: str, + model: str, + layers: int, + tp: int, + cp: int, + group: int | None, + ep: int = 1, + etp: int = 1, ) -> str: key = f"{cell}|{model}|L{layers}|tp{tp}|cp{cp}" + key += f"|ep{ep}" if ep > 1 else "" + key += f"|etp{etp}" if etp > 1 else "" return key + (f"|g{group}" if group is not None else "") @@ -72,27 +100,87 @@ def _two_gpu_cells() -> set[str]: return cells -def test_recipes_launch_exactly_the_manifest_cells() -> None: - manifest = json.loads(_MANIFEST.read_text()) +_RECIPE_CELLS = { + "dev/trainer_rank_cost_calibration_local.sh": _local_cells, + "dev/trainer_rank_cost_calibration_cp4.sky.yaml": _cp4_cells, + "dev/trainer_rank_cost_calibration_2gpu.sky.yaml": _two_gpu_cells, +} + + +def launch_cells(launch: dict) -> set[str]: + """Cells one lattice launch runs: every shape x every cell spec.""" + + assert launch["recipe"] == _LATTICE_RECIPE, launch["recipe"] + model = launch["model"] + cells: set[str] = set() + for shape in launch["shapes"]: + tp, cp, ep, etp = (list(shape) + [1, 1])[:4] + for spec in launch["cells"]: + name, *rest = spec.split(":") + layers = int(rest[0]) if rest else 0 + group = int(rest[1]) if len(rest) > 1 else None + depth = _FULL_LAYERS[model] if layers == 0 else layers + cells.add(_key(name, model, depth, tp, cp, group, ep, etp)) + return cells + + +def manifest_cells(manifest: dict) -> set[str]: + launched: set[str] = set() + for recipe in manifest.get("recipes", []): + launched |= _RECIPE_CELLS[recipe]() + for launch in manifest.get("launches", {}).values(): + launched |= launch_cells(launch) + return launched + + +@pytest.mark.parametrize("table", CALIBRATED_TABLES, ids=lambda t: t.table_id) +def test_launches_produce_exactly_the_manifest_cells(table) -> None: + manifest = json.loads(manifest_path(table.table_id).read_text()) + assert manifest["table_id"] == table.table_id expected = {cell["key"] for cell in manifest["cells"]} - launched = _local_cells() | _cp4_cells() | _two_gpu_cells() - assert launched, "no run_cell invocations parsed from the recipes" + launched = manifest_cells(manifest) + assert launched, "no launches parsed from the manifest" assert launched == expected, { - "only in recipes": sorted(launched - expected), + "only in launches": sorted(launched - expected), "only in manifest": sorted(expected - launched), } excluded = {cell["key"] for cell in manifest["excluded"]} assert excluded <= expected - assert len(expected - excluded) == 58 + assert all(cell.get("reason") for cell in manifest["excluded"]) + campaigns = {cell["campaign"] for cell in manifest["cells"]} + assert campaigns <= set(manifest.get("campaign_labels", [])) | set( + manifest.get("launches", {}) + ) -def test_manifest_campaign_labels_name_checked_in_recipes() -> None: - manifest = json.loads(_MANIFEST.read_text()) - assert {cell["campaign"] for cell in manifest["cells"]} <= { - "local", - "tr-cost-cp4", - "tr-cost-cp4-2", - "tr-cost-2gpu", - "tr-cost-2gpu-2", - "tr-cost-cp4-840", +def test_dense_table_manifest_is_the_original_58_cells_plus_blind_spot_evidence() -> ( + None +): + """The 58 fitted cells of the original recipes, plus the Qwen3-4B real-data + launch at TP1 x CP4 and TP2 x CP2 (16 cells) that documents the + context-parallel blind spot; none of those 16 is fitted.""" + + manifest = json.loads(manifest_path("dense-h2560-h200-bf16").read_text()) + expected = {cell["key"] for cell in manifest["cells"]} + excluded = {cell["key"] for cell in manifest["excluded"]} + assert len(expected - excluded) == 58 + assert excluded == launch_cells(manifest["launches"]["tr-cost-q3-4b-cp4"]) + assert len(excluded) == 16 + assert [entry["shape"] for entry in manifest["withheld"]] == [[1, 4, 1, 1]] + for recipe in manifest["recipes"]: + assert (_ROOT / recipe).is_file(), recipe + + +def test_lattice_launch_expansion() -> None: + launch = { + "recipe": _LATTICE_RECIPE, + "model": "Qwen/Qwen3.5-35B-A3B", + "shapes": [[1, 2, 2], [2, 1, 1, 1]], + "cells": ["cal-grpo-g8:0", "cal-ellavox:0:3"], + } + assert launch_cells(launch) == { + "cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2", + "cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp1|cp2|ep2|g3", + "cal-grpo-g8|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1", + "cal-ellavox|Qwen/Qwen3.5-35B-A3B|L40|tp2|cp1|g3", } diff --git a/tests/unit/test_planner_cost_profile.py b/tests/unit/test_planner_cost_profile.py index ae6841b1e..0d3e04767 100644 --- a/tests/unit/test_planner_cost_profile.py +++ b/tests/unit/test_planner_cost_profile.py @@ -2,8 +2,8 @@ A calibrated table admits a runtime when its device class, parameter dtype, model geometry (read from the Megatron config, never a model name) and -parallel shape (TP x CP x EP x ETP) all appear in the table's measured sets; -everything else keeps the version-1 score. +parallel shape (TP x CP x EP x ETP) all appear in the table's measured sets and +the pair is not withheld; everything else keeps the version-1 score. """ from __future__ import annotations @@ -57,6 +57,8 @@ def _selection( def test_measured_execution_classes_are_admitted() -> None: for geometry in (QWEN35_4B_GEOMETRY, QWEN3_4B_GEOMETRY): for shape in DENSE_H2560_TABLE.shapes: + if (geometry, shape) in DENSE_H2560_TABLE.withheld: + continue selection = _selection(geometry=geometry, shape=shape) assert selection.version == COEFFICIENT_VERSION assert selection.table_id == DENSE_H2560_TABLE.table_id @@ -71,12 +73,12 @@ def test_unmeasured_classes_fall_back_to_version_one() -> None: assert _selection(device_capability=(8, 0)).version == fallback assert _selection(device_memory_bytes=H100_MEMORY_BYTES).version == fallback assert _selection(param_dtype="torch.float16").version == fallback - # Any geometry change: a wider model (Qwen3-8B-like), a neighbouring - # width, a different FFN ratio or head geometry, GDN state shape, or MoE. + # Any geometry change: an unmeasured wider model, a neighbouring width, a + # different FFN ratio or head geometry, GDN state shape, or MoE. dense = QWEN3_4B_GEOMETRY for geometry in ( - ModelGeometry(4_096, 12_288, 32, 8, 128), - ModelGeometry(2_048, 6_144, 16, 8, 128), + ModelGeometry(6_144, 16_384, 48, 8, 128), + ModelGeometry(3_072, 8_192, 24, 8, 128), ModelGeometry(**{**dense.as_dict(), "ffn_hidden_size": 8_192}), ModelGeometry(**{**dense.as_dict(), "num_query_groups": 4}), ModelGeometry(**{**QWEN35_4B_GEOMETRY.as_dict(), "gdn_value_heads": 16}), @@ -103,6 +105,123 @@ def test_unmeasured_classes_fall_back_to_version_one() -> None: assert selection.table_id is None and selection.coefficients is None +def test_withheld_pair_falls_back_while_the_rest_of_its_table_is_admitted() -> None: + # Qwen3-4B at TP1 x CP4 is measured but withheld (the ten-term score cannot + # see the CP exchange schedule there); the GDN geometry keeps CP4 and the + # attention geometry keeps its other shapes. + cp4 = ParallelShape(tp=1, cp=4) + assert (QWEN3_4B_GEOMETRY, cp4) in DENSE_H2560_TABLE.withheld + withheld = _selection(geometry=QWEN3_4B_GEOMETRY, shape=cp4) + assert withheld.version == COEFFICIENT_VERSION_FALLBACK + assert withheld.table_id is None and withheld.coefficients is None + assert _selection(geometry=QWEN35_4B_GEOMETRY, shape=cp4).table_id == ( + DENSE_H2560_TABLE.table_id + ) + for shape in (ParallelShape(), ParallelShape(tp=1, cp=2), ParallelShape(tp=2)): + assert _selection(geometry=QWEN3_4B_GEOMETRY, shape=shape).table_id == ( + DENSE_H2560_TABLE.table_id + ), shape + + +def test_attention_classes_are_admitted_only_where_their_gates_pass() -> None: + from art.trainer_rank._planner_cost import ( + DENSE_ATTN_H2048_TABLE, + DENSE_ATTN_H4096_TABLE, + DENSE_ATTN_H5120_TABLE, + ) + + cp4 = ParallelShape(tp=1, cp=4) + tp2cp2 = ParallelShape(tp=2, cp=2) + for table in ( + DENSE_ATTN_H2048_TABLE, + DENSE_ATTN_H4096_TABLE, + DENSE_ATTN_H5120_TABLE, + ): + (geometry,) = table.geometries + for shape in table.shapes: + selection = _selection(geometry=geometry, shape=shape) + assert selection.table_id == table.table_id, (table.table_id, shape) + assert selection.coefficients is table.coefficients_milli_us + # TP1 x CP4 is admitted through the two-stage re-ranker: the table's + # shortlist score plus its certified re-ranker, never the direct table. + assert cp4 not in table.shapes and cp4 in table.reranked_shapes + reranked = _selection(geometry=geometry, shape=cp4) + assert reranked.version == COEFFICIENT_VERSION + assert reranked.table_id == table.table_id + assert reranked.reranker is table.reranker + assert reranked.coefficients is (table.reranker.shortlist_coefficients_milli_us) + # Directly scored shapes never carry a re-ranker. + assert _selection(geometry=geometry, shape=ParallelShape()).reranker is None + # A different dtype never borrows the table. + assert _selection( + geometry=geometry, shape=ParallelShape(), param_dtype="torch.float16" + ).version == (COEFFICIENT_VERSION_FALLBACK) + # TP2 x CP2 passed its gates only for the hidden-5,120 class. + assert tp2cp2 in DENSE_ATTN_H5120_TABLE.shapes + assert tp2cp2 not in DENSE_ATTN_H4096_TABLE.shapes + assert tp2cp2 not in DENSE_ATTN_H2048_TABLE.shapes + + +def test_attention_moe_class_admits_only_its_certified_shapes() -> None: + from art.trainer_rank._planner_cost import ( + ATTN_MOE_H2048_GEOMETRY, + ATTN_MOE_H2048_TABLE, + GDN_MOE_H2048_GEOMETRY, + ) + + geometry = ATTN_MOE_H2048_GEOMETRY + for shape in ATTN_MOE_H2048_TABLE.shapes: + selection = _selection(geometry=geometry, shape=shape) + assert selection.table_id == ATTN_MOE_H2048_TABLE.table_id, shape + assert selection.reranker is None + # TP1 x CP4 is re-ranked at every measured expert parallelism. + for ep in (1, 2, 4): + reranked = _selection(geometry=geometry, shape=ParallelShape(tp=1, cp=4, ep=ep)) + assert reranked.reranker is ATTN_MOE_H2048_TABLE.reranker, ep + # Unmeasured shapes keep the version-1 score. + for shape in ( + ParallelShape(tp=2, cp=2), + ParallelShape(tp=1, cp=8), + ParallelShape(tp=1, cp=4, ep=8), + ParallelShape(tp=2, cp=1, ep=2, etp=2), + ): + assert _selection(geometry=geometry, shape=shape).version == ( + COEFFICIENT_VERSION_FALLBACK + ), shape + # The GDN MoE class and this attention MoE class never borrow each other. + assert ( + _selection( + geometry=GDN_MOE_H2048_GEOMETRY, shape=ParallelShape(tp=2, cp=1, ep=2) + ).table_id + != ATTN_MOE_H2048_TABLE.table_id + ) + + +def test_dense_gdn_h5120_class_admits_its_four_measured_shapes() -> None: + from art.trainer_rank._planner_cost import ( + DENSE_ATTN_H5120_TABLE, + DENSE_GDN_H5120_TABLE, + QWEN3_14B_GEOMETRY, + QWEN35_27B_GEOMETRY, + ) + + assert DENSE_GDN_H5120_TABLE.reranker is None + for shape in DENSE_GDN_H5120_TABLE.shapes: + selection = _selection(geometry=QWEN35_27B_GEOMETRY, shape=shape) + assert selection.table_id == DENSE_GDN_H5120_TABLE.table_id, shape + # CP4 is scored directly for this GDN class; TP2 x CP2 keeps version 1. + assert ParallelShape(tp=1, cp=4) in DENSE_GDN_H5120_TABLE.shapes + assert _selection( + geometry=QWEN35_27B_GEOMETRY, shape=ParallelShape(tp=2, cp=2) + ).version == (COEFFICIENT_VERSION_FALLBACK) + # Same hidden size as the Qwen3-14B attention class, different geometry: + # the two tables never borrow each other. + assert QWEN35_27B_GEOMETRY.hidden_size == QWEN3_14B_GEOMETRY.hidden_size + assert _selection( + geometry=QWEN3_14B_GEOMETRY, shape=ParallelShape(tp=1, cp=1) + ).table_id == (DENSE_ATTN_H5120_TABLE.table_id) + + def test_cpu_only_planning_uses_the_default_table() -> None: # No CUDA device (unit tests): the calibrated domain describes GPU # execution, so the device is not a reason to fall back. @@ -227,3 +346,37 @@ def test_scores_follow_the_selected_table() -> None: # Four 2,100-token sequences sharing a 2,000-token prefix: the selected # layout packs fewer tokens than the unshared 8,400 and more than the prefix. assert 2_000 < selected.layout.packed_tokens < 8_400 + + +def test_moe_class_is_admitted_only_at_its_measured_shapes() -> None: + from art.trainer_rank._planner_cost import ( + GDN_MOE_H2048_GEOMETRY, + GDN_MOE_H2048_TABLE, + ) + + for shape in GDN_MOE_H2048_TABLE.shapes: + selection = _selection(geometry=GDN_MOE_H2048_GEOMETRY, shape=shape) + assert selection.table_id == GDN_MOE_H2048_TABLE.table_id, shape + assert ParallelShape(tp=1, cp=4, ep=4) in GDN_MOE_H2048_TABLE.shapes + # Unmeasured shapes (TP2 x CP2, EP1 at CP8, expert TP) and the dense table's + # shapes with this geometry under a different dtype fall back. + for shape in ( + ParallelShape(tp=2, cp=2), + ParallelShape(tp=1, cp=8), + ParallelShape(tp=2, cp=1, ep=2, etp=2), + ): + assert _selection(geometry=GDN_MOE_H2048_GEOMETRY, shape=shape).version == ( + COEFFICIENT_VERSION_FALLBACK + ), shape + assert ( + _selection( + geometry=GDN_MOE_H2048_GEOMETRY, + shape=ParallelShape(), + param_dtype="torch.float16", + ).version + == COEFFICIENT_VERSION_FALLBACK + ) + # The dense geometries never borrow the MoE table and vice versa. + assert _selection( + geometry=QWEN35_4B_GEOMETRY, shape=ParallelShape(tp=1, cp=4, ep=4) + ).version == (COEFFICIENT_VERSION_FALLBACK) diff --git a/tests/unit/test_planner_rerank.py b/tests/unit/test_planner_rerank.py new file mode 100644 index 000000000..e25f0bcdb --- /dev/null +++ b/tests/unit/test_planner_rerank.py @@ -0,0 +1,232 @@ +"""Two-stage layout selection on re-ranked parallel shapes. + +On a table's ``reranked_shapes`` the ten-term score only shortlists: the +cheapest ``shortlist_size`` layouts plus the incumbent anchor are priced by +the context-parallel plan each produces (remote wave count, largest per-rank +token load), and the lowest second-stage score is selected. These tests use +a synthetic re-ranker and a synthetic plan structure; the certified +coefficients are bound to their certificates elsewhere. +""" + +from __future__ import annotations + +import pytest +import torch + +from art.trainer_rank._planner_cost import ( + COEFFICIENT_VERSION, + COEFFICIENTS_MILLI_US, + DENSE_H2560_TABLE, + H200_CLASS, + QWEN3_4B_GEOMETRY, + CalibratedTable, + DeviceClass, + ParallelShape, + ReRanker, + prefix_tree_layout_score, +) +from art.trainer_rank._prefix_tree_planner import ( + CanonicalPrefixTree, + LayoutCandidate, + build_canonical_prefix_tree, + prefix_tree_layout_candidates, + select_prefix_tree_layout, +) + + +def _tree() -> CanonicalPrefixTree: + # A two-level prefix tree: a shared 2,000-token prompt, two 600-token + # branches, two 200-token leaves each, so the mandatory family has + # several distinct layouts (no sharing, depth one, depth two, spans). + prompt = list(range(1, 2_001)) + branches = [list(range(3_000 + b * 1_000, 3_600 + b * 1_000)) for b in range(2)] + rows = [ + torch.tensor( + prompt + + branches[b] + + list(range(9_000 + (2 * b + leaf) * 300, 9_200 + (2 * b + leaf) * 300)) + ) + for b in range(2) + for leaf in range(2) + ] + return build_canonical_prefix_tree(tuple(rows)) + + +def _reranker( + *, + shortlist_size: int = 3, + wave_per_layer_milli_us: int = 1_000_000, + max_rank_token_per_layer_milli_us: int = 1_000, +) -> ReRanker: + return ReRanker( + shortlist_size=shortlist_size, + incumbent="depth_one", + shortlist_coefficients_milli_us=COEFFICIENTS_MILLI_US, + wave_per_layer_milli_us=wave_per_layer_milli_us, + max_rank_token_per_layer_milli_us=max_rank_token_per_layer_milli_us, + ) + + +def _cheap_order( + tree: CanonicalPrefixTree, cp_size: int = 4, layers: int = 8 +) -> list[LayoutCandidate]: + scored = sorted( + prefix_tree_layout_candidates(tree), + key=lambda c: prefix_tree_layout_score( + c.layout, cp_size=cp_size, layers=layers, uses_gdn=False + ), + ) + return scored + + +def test_reranker_scores_waves_and_load_per_layer() -> None: + reranker = _reranker(wave_per_layer_milli_us=7, max_rank_token_per_layer_milli_us=2) + assert reranker.score(layers=10, wave_count=3, max_rank_tokens=100) == 10 * ( + 21 + 200 + ) + + +def _table( + *, + reranked_shapes: tuple[ParallelShape, ...] = (), + reranker: ReRanker | None = None, +) -> CalibratedTable: + return CalibratedTable( + table_id="t", + coefficients_milli_us=COEFFICIENTS_MILLI_US, + device_classes=(H200_CLASS,), + param_dtypes=("torch.bfloat16",), + geometries=(QWEN3_4B_GEOMETRY,), + shapes=(ParallelShape(),), + reranked_shapes=reranked_shapes, + reranker=reranker, + ) + + +def test_table_validation_ties_reranked_shapes_to_a_reranker() -> None: + cp4 = ParallelShape(tp=1, cp=4) + with pytest.raises(ValueError): + _table(reranked_shapes=(cp4,)) + with pytest.raises(ValueError): + _table(reranker=_reranker()) + with pytest.raises(ValueError): + _table(reranked_shapes=(ParallelShape(),), reranker=_reranker()) + table = _table(reranked_shapes=(cp4,), reranker=_reranker()) + device = DeviceClass((9, 0), "hbm-141g") + dtype = "torch.bfloat16" + assert table.admits( + device=device, + param_dtype=dtype, + geometry=QWEN3_4B_GEOMETRY, + shape=ParallelShape(), + ) + assert not table.admits( + device=device, param_dtype=dtype, geometry=QWEN3_4B_GEOMETRY, shape=cp4 + ) + assert table.reranks( + device=device, param_dtype=dtype, geometry=QWEN3_4B_GEOMETRY, shape=cp4 + ) + assert not table.reranks( + device=device, + param_dtype=dtype, + geometry=QWEN3_4B_GEOMETRY, + shape=ParallelShape(tp=1, cp=2), + ) + assert not table.reranks( + device=device, + param_dtype="torch.float16", + geometry=QWEN3_4B_GEOMETRY, + shape=cp4, + ) + assert DENSE_H2560_TABLE.reranker is None and not DENSE_H2560_TABLE.reranked_shapes + + +def test_without_a_reranker_selection_is_the_cheap_incumbent() -> None: + tree = _tree() + plain = select_prefix_tree_layout( + tree, cp_size=4, layers=8, uses_gdn=False, refinement_work_budget=0 + ) + assert plain.layout == _cheap_order(tree)[0].layout + + +def test_reranking_picks_the_lowest_second_stage_score_in_the_shortlist() -> None: + tree = _tree() + order = _cheap_order(tree) + assert len(order) >= 4 + # The plan structure makes the cheap order's third layout the clear winner + # (one wave, smallest load) and the cheap winner the worst. + structure = {c.layout: (2, 5_000) for c in order} + structure[order[2].layout] = (1, 1_000) + structure[order[0].layout] = (4, 9_000) + selected = select_prefix_tree_layout( + tree, + cp_size=4, + layers=8, + uses_gdn=False, + refinement_work_budget=0, + reranker=_reranker(shortlist_size=3), + plan_structure=lambda layout: structure[layout], + ) + assert selected.layout == order[2].layout + # Outside the shortlist a layout is never selected however good its plan. + structure[order[-1].layout] = (0, 1) + if order[-1] not in order[:3] and "depth_one" not in order[-1].labels: + selected = select_prefix_tree_layout( + tree, + cp_size=4, + layers=8, + uses_gdn=False, + refinement_work_budget=0, + reranker=_reranker(shortlist_size=3), + plan_structure=lambda layout: structure[layout], + ) + assert selected.layout == order[2].layout + + +def test_ties_keep_the_cheaper_layout_and_the_incumbent_is_always_considered() -> None: + tree = _tree() + order = _cheap_order(tree) + # Identical plan structure everywhere: the re-ranker must not disturb the + # cheap order. + flat = lambda layout: (1, 4_000) # noqa: E731 + selected = select_prefix_tree_layout( + tree, + cp_size=4, + layers=8, + uses_gdn=False, + refinement_work_budget=0, + reranker=_reranker(shortlist_size=2), + plan_structure=flat, + ) + assert selected.layout == order[0].layout + # A shortlist of one still considers the depth-one anchor. + depth_one = next(c for c in order if "depth_one" in c.labels) + if depth_one is not order[0]: + structure = {c.layout: (3, 9_000) for c in order} + structure[depth_one.layout] = (1, 1_000) + selected = select_prefix_tree_layout( + tree, + cp_size=4, + layers=8, + uses_gdn=False, + refinement_work_budget=0, + reranker=_reranker(shortlist_size=1), + plan_structure=lambda layout: structure[layout], + ) + assert selected.layout == depth_one.layout + + +def test_reranked_selection_requires_a_plan_structure() -> None: + with pytest.raises(ValueError): + select_prefix_tree_layout( + _tree(), + cp_size=4, + layers=8, + uses_gdn=False, + refinement_work_budget=0, + reranker=_reranker(), + ) + + +def test_selection_version_is_unchanged_by_the_reranker_plumbing() -> None: + assert COEFFICIENT_VERSION == 2 diff --git a/tests/unit/test_trainer_rank_cost_fit.py b/tests/unit/test_trainer_rank_cost_fit.py index e7950c8a8..ccfb8e4e6 100644 --- a/tests/unit/test_trainer_rank_cost_fit.py +++ b/tests/unit/test_trainer_rank_cost_fit.py @@ -175,3 +175,81 @@ def candidates(cell: str, model: str, times: dict[str, float]) -> list: ) assert report["M1"]["max_regret_pct"] == 0.0 and not report["M1"]["gate_problems"] assert report["M2"]["max_regret_pct"] > 10.0 and report["M2"]["gate_problems"] + + +def test_integerize_keeps_plain_rounding_when_it_preserves_rankings() -> None: + import numpy as np + + cells = [ + fit.Candidate( + "a|M|L2|tp1|cp1", + label, + { + "packed_tokens": tokens, + "segment_count": 1, + "max_depth": 1, + "segments_below": (), + }, + { + "layers": 2.0, + "gdn_layers": 0.0, + "tp": 1.0, + "cp": 1.0, + "ep": 1.0, + "etp": 1.0, + "uses_gdn": 0.0, + }, + ms, + 8, + 0.5, + ) + for label, (ms, tokens) in {"x": (100.0, 1000), "y": (120.0, 1500)}.items() + ] + assert fit.integerize(cells, ("token_per_rank",), np.asarray([2.0004])) == { + "token_per_rank": 2000 + } + + +def test_integerize_never_loses_to_plain_rounding() -> None: + import numpy as np + + def candidates(cell: str, times: dict[str, tuple[float, int]]) -> list: + return [ + fit.Candidate( + cell, + label, + { + "packed_tokens": tokens, + "segment_count": 1, + "max_depth": 1, + "segments_below": (), + }, + { + "layers": 2.0, + "gdn_layers": 0.0, + "tp": 1.0, + "cp": 1.0, + "ep": 1.0, + "etp": 1.0, + "uses_gdn": 0.0, + }, + ms, + 8, + 0.5, + ) + for label, (ms, tokens) in times.items() + ] + + cells = candidates( + "a|M|L2|tp1|cp1", {"x": (100.0, 1000), "y": (110.0, 1100)} + ) + candidates("b|M|L2|tp1|cp1", {"x": (100.0, 1000), "y": (99.0, 1001)}) + terms = ("token_per_rank",) + beta = np.asarray([0.0004]) # rounds to 0 milli-us: every cell would tie + table = fit.integerize(cells, terms, beta) + assert all(isinstance(v, int) for v in table.values()) + matrix = fit.term_matrix(cells, terms) + rounded = np.asarray([0.0]) + refined = np.asarray([table["token_per_rank"] / 1_000.0]) + assert fit.selection_loss(cells, matrix @ refined) <= fit.selection_loss( + cells, matrix @ rounded + ) diff --git a/tests/unit/test_trainer_rank_validation.py b/tests/unit/test_trainer_rank_validation.py index 4eb3f362f..5928ddf94 100644 --- a/tests/unit/test_trainer_rank_validation.py +++ b/tests/unit/test_trainer_rank_validation.py @@ -620,10 +620,13 @@ def test_hybridep_uses_maximum_cp_model_rows( configured = trainer._configure_hybridep((batch, short_batch), topology=topology) + # The reserved capacity covers the largest planned rows per rank (13, from + # the longer batch) as well as the near-balanced extent of the sequence. assert calls == { "capacity": { "packed_sequence_length": 9, "context_parallel_size": 4, + "required_capacity": 13, }, } assert configured == ((13, 11), 13)