feat: upgrade vllm from 0.25.1 to 0.27.1 for CUDA and ROCm backends - #251
Open
axsapronov wants to merge 3 commits into
Open
feat: upgrade vllm from 0.25.1 to 0.27.1 for CUDA and ROCm backends#251axsapronov wants to merge 3 commits into
axsapronov wants to merge 3 commits into
Conversation
Key dependency changes in upstream vllm 0.27.1:
- PyTorch: 2.11.0 → 2.13.0 (breaking change)
- Torchvision: 0.22.0 → 0.28.0
- Triton: 3.1.0 → 3.7.1
- FlashInfer: 0.6.x → 0.6.16.post3
- NCCL: 2.21.x → 2.29.7
- VLLM_OMNI_COMMIT → d77a35a3 (v0.27.0rc1-25-gd77a35a3, rebased onto vLLM 0.27.x)
Updated all patches to match v0.27.1 source context:
- 001_wrong_dp_ray.patch: get_open_port() structure unchanged, context
adjusted for new _get_open_port() signature (start_port, max_attempts)
- 002_shm_broadcast_port_race.patch: TOCTOU race condition still present
in shm_broadcast.py, line numbers shifted due to SPDX headers
- vllm_omni/001_wrong_patch.patch: patch.py grew significantly (300 → 573
lines) with additional patches (NVFP4 NaN clamp, CuMemAllocator fix,
inductor divisibility), but sys.modules monkey-patching loop still lacks
--omni guard — context rewritten for new line numbers
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request updates the vLLM version to 0.27.1 across CUDA and ROCm backends, including updates to Dockerfiles, runner configurations, and patches. The reviewer feedback suggests adding a final newline to the modified patch files and restructuring the project to share common patches between CUDA and ROCm to reduce duplication.
…llm 0.27.1 Test fixtures had hardcoded expected values for CUDA and ROCm backends that did not include the new vllm 0.27.1 entries. Regenerated the fixture from actual runner catalog data so all 8 backend test cases pass.
- Add final newline to patch files (convention for text files) - Extract common vllm and vllm_omni patches to pack/common/patches/ - Replace cuda and rocm patches with symlinks to common patches to reduce duplication
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Upgrade vllm from 0.25.1 to 0.27.1 across CUDA 13.0, CUDA 12.9, and ROCm 7.2 backends.
Key upstream dependency changes in vllm 0.27.1
Changes
pack/cuda/Dockerfile.vllm,pack/rocm/Dockerfile.vllm): UpdatedVLLM_VERSION,VLLM_BASE_IMAGE,VLLM_TORCH_VERSION, andVLLM_OMNI_COMMITpack/matrix.yaml: Updated all three vllm build rules (CUDA 13.0, CUDA 12.9, ROCm 7.2)README.md: Added v0.27.1 as polished version for CUDA 13.0, CUDA 12.9, and ROCm 7.2gpustack_runner/runner.py.json: Added 5 new runner entries for vllm 0.27.1d3c47ef(v0.25.x rebase) tod77a35a3(v0.27.0rc1-25-gd77a35a3, HEAD main)Patches verified
All three patches (
001_wrong_dp_ray.patch,002_shm_broadcast_port_race.patch,001_wrong_patch.patch) were tested withpatch --dry-runagainst v0.27.1 source and pass successfully.