From 9c3e8b02955ccfc0a3f7e3974caa951708076e5a Mon Sep 17 00:00:00 2001 From: twilson63 Date: Wed, 18 Nov 2026 17:04:09 -0500 Subject: [PATCH] =?UTF-8?q?Fix:=20Ornith=2035B=20A3B=20OV=20results=20?= =?UTF-8?q?=E2=80=94=20quant=20is=20int8-asym,=20not=20int4?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Community reviewer @Ratgor flagged the HF artifact (Ornith-1.5-35B-A3B-INT4-ov discussion #1): the language model XML carries 1082 U8 tensors + zero_points and zero i4/u4 tensors; 31.9 GB bin for 35B params = 0.91 B/param. Same optimum-intel 2.2.0 + NNCF 3.4 data-aware fallback as the Mellum fix in #119. quant field corrected; evidence note added to notes. Flags text kept under the 200-char limit. --- results/twilson63/ornith-35b-b70-cascadia.json | 8 ++++---- results/twilson63/ornith-35b-b70.json | 8 ++++---- 2 files changed, 8 insertions(+), 8 deletions(-) diff --git a/results/twilson63/ornith-35b-b70-cascadia.json b/results/twilson63/ornith-35b-b70-cascadia.json index 5388c98..0e0611a 100644 --- a/results/twilson63/ornith-35b-b70-cascadia.json +++ b/results/twilson63/ornith-35b-b70-cascadia.json @@ -4,13 +4,13 @@ "component": "intel-arc-pro-b70", "componentQuantity": 1, "model": "ornith-1-5-35b-a3b", - "quant": "int4", + "quant": "int8", "runtime": "cascadia", "runtimeVersion": "0.2.4", - "runtimeFlags": "prebuilt bundle, ov-genai engine, device GPU, batch 1, OpenAI API", + "runtimeFlags": "prebuilt bundle, ov-genai engine, device GPU, batch 1, OpenAI API \u2014 quant correction 2026-11-18: weights are INT8 asymmetric (U8 + zero-points, NNCF fallback from requested int4-sym; verified via", "decodeTps": 43.1, "contextLength": 143, "batchSize": 1, "runDate": "2026-09-28", - "notes": "ReBAR on, 32GB BAR. 2 runs: 128 tokens in 3.181s/2.753s wall. End-to-end wall incl. prefill; TTFT not exposed by API." -} + "notes": "ReBAR on, 32GB BAR. 2 runs: 128 tokens in 3.181s/2.753s wall. End-to-end wall incl. prefill; TTFT not exposed by API. | Quant correction 2026-11-18: weights are INT8 asymmetric (U8 + zero-points, NNCF fallback from requested int4-sym; verified via XML, zero i4/u4 tensors)." +} \ No newline at end of file diff --git a/results/twilson63/ornith-35b-b70.json b/results/twilson63/ornith-35b-b70.json index 1ea13ff..2c358b2 100644 --- a/results/twilson63/ornith-35b-b70.json +++ b/results/twilson63/ornith-35b-b70.json @@ -4,15 +4,15 @@ "component": "intel-arc-pro-b70", "componentQuantity": 1, "model": "ornith-1-5-35b-a3b", - "quant": "int4", + "quant": "int8", "runtime": "openvino-genai", "runtimeVersion": "2026.4.0.0-3407-7ea2546852a", - "runtimeFlags": "OpenVINO 2026.4.0, VLMPipeline text-only, default GenAI pipeline, batch 1", + "runtimeFlags": "OpenVINO 2026.4.0, VLMPipeline text-only, default GenAI pipeline, batch 1 \u2014 quant correction 2026-11-18: weights are INT8 asymmetric (U8 + zero-points, NNCF fallback from requested int4-sym; verified", "decodeTps": 72.2, "promptTps": 58.1, "ttftMs": 442.6, "contextLength": 153, "batchSize": 1, "runDate": "2026-09-28", - "notes": "ReBAR on, 32GB BAR. 3-run mean; decode_tps runs: 67.6/75.4/73.6. Community OV convert vonvonhero/Ornith-1.5-35B-A3B-Uncensored-OpenVINO-INT4, run text-only via VLMPipeline. prompt_tps = input_tokens/TTFT." -} + "notes": "ReBAR on, 32GB BAR. 3-run mean; decode_tps runs: 67.6/75.4/73.6. Community OV convert vonvonhero/Ornith-1.5-35B-A3B-Uncensored-OpenVINO-INT4, run text-only via VLMPipeline. prompt_tps = input_tokens/TTFT. | Quant correction 2026-11-18: weights are INT8 asymmetric (U8 + zero-points, NNCF fallback from requested int4-sym; verified via XML, zero i4/u4 tensors)." +} \ No newline at end of file