Repository navigation

CI follow-up: cache source-built wheels and extensions, bump CI containers, trim slow GPU tests - #2836

Merged

kevalmorabia97 merged 38 commits into

mainfrom

kmorabia/ci-gpu-followup

Oct 20, 2026

Merged

CI follow-up: cache source-built wheels and extensions, bump CI containers, trim slow GPU tests#2836
kevalmorabia97 merged 38 commits into
mainfrom
kmorabia/ci-gpu-followup

Commits on Oct 16, 2026
- Worker pool: report a failed rank before its teardown; move the helper test to unit
  
  kevalmorabia97
  and
  claude
  committed
  b7b7657
- CI: cache source-built wheels in the gpu job and bump the CI containers
  
  kevalmorabia97
  and
  claude
  committed
  e5d149e
- Remove the duplicate gpu_megatron copy of the CUDA extension tests
  
  kevalmorabia97
  and
  claude
  committed
  4aedd5f
- gpu job: save the wheel cache even when the tests fail or time out
  
  kevalmorabia97
  and
  claude
  committed
  95b4913
- CI: save the CUDA extension cache even when the job fails or times out
  
  kevalmorabia97
  and
  claude
  committed
  7e8cfd1
- EAGLE: compile _eagle_forward without CUDA graphs
  
  kevalmorabia97
  and
  claude
  committed
  8c5fadc
- gpu job: build mamba_ssm and causal-conv1d as C++20
  
  kevalmorabia97
  and
  claude
  committed
  613e40e
- NVFP4: convert cutlass-swizzled block scales in ModelOpt instead of importing them from TensorRT-LLM
  
  kevalmorabia97
  and
  claude
  committed
  ae6f163
- Keep vLLM at v0.30.0: v0.31.0 breaks the vLLM plugin
  
  kevalmorabia97
  and
  claude
  committed
  e0dafde
- Address review: wheelhouse path, complete-only cache save, wheel lookup, force-build, teardown error, test margins
  
  kevalmorabia97
  and
  claude
  committed
  27235c7
- trtllm examples: install fla-core for TensorRT-LLM rc29
  
  kevalmorabia97
  and
  claude
  committed
  db68967
- Pin TensorRT-LLM to rc23; build the conv3d implicit-GEMM extension without a hardcoded C++17
  
  kevalmorabia97
  and
  claude
  committed
  4ffb17e
- persistent_materialization: skip FSDP2 unshard only for the materialized layer
  
  kevalmorabia97
  and
  claude
  committed
  a04bd58
- noxfile: drop the stale cppimport note
  
  kevalmorabia97
  and
  claude
  committed
  878ffc5
- Merge remote-tracking branch 'origin/main' into kmorabia/ci-gpu-followup
  kevalmorabia97
  committed
  6253454
- Address review: retire a pool after any failed job; restore nested FSDP unshard/reshard overrides
  
  kevalmorabia97
  and
  claude
  committed
  50615b3
- NVFP4: return a contiguous scale from cutlass_fp4_scale_to_modelopt_fp4_scale
  
  kevalmorabia97
  and
  claude
  committed
  635c761
- Keep TensorRT-LLM at rc20
  
  kevalmorabia97
  and
  claude
  committed
  f181890
- gpu_megatron: destroy each test's process groups in the worker teardown
  
  kevalmorabia97
  and
  claude
  committed
  9d3d714
Commits on Oct 17, 2026
- Merge remote-tracking branch 'origin/main' into kmorabia/ci-gpu-followup
  kevalmorabia97
  committed
  9d9ab38
- Use a fixed tile config for the NVFP4 FP8 sweep kernel; drop unneeded xfails
  
  kevalmorabia97
  and
  claude
  committed
  7fc0595
- Test the FSDP2 unshard/reshard override with a shared param group and nesting
  
  kevalmorabia97
  and
  claude
  committed
  528c39e
- Key the torch-extension cache on the C++/CUDA sources only
  
  kevalmorabia97
  and
  claude
  committed
  1203fcd
- Drop the vLLM v0.20.0 gpu_vllm job
  
  kevalmorabia97
  and
  claude
  committed
  0ef079a
- Only skip FSDP unshard/reshard when the materialized layer is FSDP-wrapped
  
  kevalmorabia97
  and
  claude
  committed
  f222d46
- Raise gpu_vllm timeout to 20 min
  
  kevalmorabia97
  and
  claude
  committed
  f086740
- Run only 3 puzzletron models in CI and raise test_mamba_search_space timeout
  
  kevalmorabia97
  and
  claude
  committed
  996f48e
- Key the extension cache per job and skip saving an empty one
  
  kevalmorabia97
  and
  claude
  committed
  f1311ea
Commits on Oct 20, 2026
- Merge remote-tracking branch 'origin/main' into kmorabia/ci-gpu-followup
  kevalmorabia97
  committed
  39f45bd
- Fix the new 2-GPU nightly failures: vLLM /dev/shm and the packed distill test's head_dim
  
  kevalmorabia97
  and
  claude
  committed
  8f58d40
- Raise the gpu_vllm timeout to 30 min for the 2-GPU nightly
  
  kevalmorabia97
  and
  claude
  committed
  f25ff98
- Raise the gpu_megatron timeout to 75 min for the 2-GPU nightly
  
  kevalmorabia97
  and
  claude
  committed
  165cdd6
- Drop the comments on the gpu_vllm and gpu_megatron timeouts
  
  kevalmorabia97
  and
  claude
  committed
  378b5b6
- Fix unit CI: pin leptonai<0.47 for the launcher; halve the DFlash2 overfit test
  
  kevalmorabia97
  and
  claude
  committed
  49ab19f
- Skip two more Megatron quantization tests hit by the known Blackwell illegal-memory-access flake
  
  kevalmorabia97
  and
  claude
  committed
  3ae2c85
- Merge branch 'main' into kmorabia/ci-gpu-followup
  kevalmorabia97
  committed
  d704983
- gpu_megatron: stop destroying NCCL groups in the worker teardown; drop the new Blackwell skips
  
  kevalmorabia97
  and
  claude
  committed
  6bc0119
- hf_eagle: note that CUDA-graph capture is currently off where rotary_emb init is hoisted
  
  kevalmorabia97
  and
  claude
  committed
  cb1cac9

Commits

Commits on Oct 16, 2026

Commits on Oct 17, 2026

Commits on Oct 20, 2026