transformers

85577dbb - CI: add vllm-test-init and vllm-test-transformers jobs on dedicated runners (#47934)

CI: add vllm-test-init and vllm-test-transformers jobs on dedicated runners (#47934) * Test vLLM CI on larger runner with test_init and test_transformers enabled Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Trigger tests_processors CI run * Split test_initialization into 8-shard matrix job; add HF_HOME env - New `vllm-test-init` matrix job runs test_initialization.py across 8 parallel shards using pytest-shard (--num-shards=8 --shard-id) - Disable the sequential test_initialization step in the `vllm` job - Add HF_HOME: /mnt/cache at workflow level Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Trigger vLLM CI run * CI: isolate JambaForCausalLM test (1 shard) to debug mamba_mixer IndexError Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: switch isolated test to FalconMambaForCausalLM (same mamba_mixer IndexError) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: patch mamba_mixer conv1d weight.view to handle 2D dummy-loaded weight Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: patch mamba_mixer out_proj transpose contiguous for CPU onednn_mm Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: fix mamba_mixer sed patch (preserve closing paren of out_proj call) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: add detailed vLLM upstream bug comments for mamba_mixer CPU patches Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: add GPU job to verify mamba_mixer bugs are CPU-specific Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: fix GPU test-init job (no curl, shell bash, ipc host, pip not uv) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: install vLLM from CUDA 13.0 nightly wheel for GPU test job Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: unset VLLM_TARGET_DEVICE for GPU job, add shell bash to pytest step Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: use PyPI vllm for GPU job (nightly wheel has broken _C_stable_libtorch) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: disable GPU test-init job (vllm._C_stable_libtorch ABI mismatch) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: run test_can_initialize_small_subset (8 shards), comment out mamba patches Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: reduce vllm-test-init to 4 shards (small subset has only 12 tests) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: skip Gemma3nForCausalLM in vllm-test-init (hangs on shm_broadcast) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: disable vllm-test-init job (Gemma3nForCausalLM hangs on shm_broadcast) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: add vllm-test-transformers job on dedicated aws-m8i-8xl-cache runner Previously ran as a step in the main vllm job and hit OOM because RAM wasn't released between steps. Now runs as a dedicated job with a fresh runner (same aws-m8i-8xl-cache group, 128 GiB RAM). Mamba patches enabled since the full transformers test suite includes Mamba models. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: add sharding (4 shards) to vllm-test-transformers job Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: add sentence-transformers to vllm-test-transformers install Required by tests/conftest.py for test_pooling[TransformersEmbeddingModel]. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: add comment to vllm-test-transformers gpu_memory patch step Previously ran as a step in the vllm job but hit OOM on the 32 GiB runner. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: re-enable vllm-test-init, vllm-multimodal-processing, and inline test steps - vllm-test-init: re-enabled (Gemma3nForCausalLM still deselected) - vllm-multimodal-processing: re-enabled - test_mapping, example chat/vision/audio steps: re-enabled Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: revert vllm job to 2xl runner, drop gpu_memory patches, restore if: always() - vllm job: back to aws-m8i-2xl-cache (heavy tests moved to dedicated jobs) - Remove gpu_memory_utilization patches from vllm job (try without them on 2xl) - Restore if: always() on test_mapping and example steps Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: move vllm and multimodal jobs to 8xl runner, restore gpu_memory patches - vllm job: back to aws-m8i-8xl-cache (examples OOM on 2xl, read-only FS issue) - vllm-multimodal-processing: 2xl → 8xl (same read-only FS issue on 2xl) - Restore gpu_memory_utilization patches in vllm job (still needed on 8xl due to container memory ceiling vs. vLLM's startup RAM check) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: document why all vllm jobs use aws-m8i-8xl-cache runner The 8xl runner is needed for both sufficient RAM and a writable shared cache mount — the 2xl runner has the same mount read-only. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: revert qwen3_asr test change, clean up comments - Revert tests/models/qwen3_asr/test_processing_qwen3_asr.py to main - gpu_memory patch comment: say probably not needed on 8xl but not verified - Remove "(fast, no sharding needed)" from vllm-test-init comment - Remove job-level comment from vllm-test-transformers Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: clarify why vllm-test-init uses 8xl runner Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * CI: remove stale comments from gpu_memory patch steps Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>