docker: upgrade base to vLLM 0.23.0, remove CUDA 13 build path - #253
Merged
Conversation
- Base image: v0.22.0-cu129-ubuntu2404 → v0.23.0-cu129-ubuntu2404 - Remove `ENABLE_CUDA_13` ARG and all conditional cu13 blocks: - cu13 apt dev headers (libcublas-dev-13-0, cuda-nvrtc-dev-13-0, etc.) - TE source build (cu13 wheel didn't exist; cu129 wheel works on arm64) - fzyzcjy triton source build (cu13 specific) - TMS_CUDA_MAJOR export (no longer needed) - Simplify cublas-dev to unconditional libcublas-dev-12-9 - Simplify TE install to wheel-only - justfile: remove `build-cu13` target and cu13 tag scheme - vllm.patch: adapt line numbers for 0.23.0 (776/1896 vs 750/1844), preserve `with self.log_iteration_details(None):` wrapper cu129 nvcc already supports sm100/sm120 (Blackwell), so cu13 build path was unnecessary — it caused build failures on gb300 (cu13 apt packages hijacked /etc/alternatives/cuda, breaking TE CMake). Tested: built successfully on gb200 (arm64), h200 (x86), gb300 (arm64). All three confirmed vLLM 0.23.0. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: aoshen02 <aoshen@inferact.ai>
![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 upgrades the base Docker image to vllm/vllm-openai:v0.23.0-cu129-ubuntu2404 and removes the experimental CUDA 13 build paths, simplifying the Dockerfile, justfile, and updating the vLLM patch file accordingly. The review feedback suggests optimizing the Dockerfile by adding --no-install-recommends to the apt-get install command and --no-cache-dir to the pip install command to minimize the final image size.
| else apt-get install -y libcublas-dev-12-9; fi && \ | ||
| rm -rf /var/lib/apt/lists/* | ||
| # cublas dev header for TE CMake (arm64 base ships runtime .so but not the header). | ||
| RUN apt-get update && apt-get install -y libcublas-dev-12-9 && rm -rf /var/lib/apt/lists/* |
There was a problem hiding this comment.
To keep the Docker image size as small as possible, it is recommended to use the --no-install-recommends flag with apt-get install. This prevents the installation of recommended but non-essential packages.
RUN apt-get update && apt-get install -y --no-install-recommends libcublas-dev-12-9 && rm -rf /var/lib/apt/lists/*
| else \ | ||
| pip -v install --no-build-isolation "transformer_engine[pytorch]==2.10.0"; \ | ||
| fi | ||
| RUN pip -v install --no-build-isolation "transformer_engine[pytorch]==2.10.0" |
There was a problem hiding this comment.
It is recommended to use the --no-cache-dir flag with pip install to prevent caching of wheel files and package downloads, which helps minimize the Docker image size.
RUN pip -v install --no-cache-dir --no-build-isolation "transformer_engine[pytorch]==2.10.0"
…l_utils weight-reload fix (#45989) v0.23.0 already includes the sleep/scheduler guard from #44483, so the core.py patch is no longer needed. Replace it with the FP8+DeepEP weight-reload fix (vllm-project/vllm#45989): snapshot max_num_batched_tokens from FusedMoEConfig instead of calling get_current_vllm_config() during layerwise reload. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Ao Shen <aoshen@inferact.ai> Signed-off-by: aoshen02 <aoshen@inferact.ai>
aoshen02
force-pushed
the
docker/upgrade-vllm-v023
branch
2 times, most recently
from
June 18, 2026 09:20
cfa977a to
d41f0aa
Compare
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
2 participants
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
v0.22.0-cu129-ubuntu2404tov0.23.0-cu129-ubuntu2404ENABLE_CUDA_13conditional build paths (cu13 apt headers, TE source build, triton source build, TMS_CUDA_MAJOR)build-cu13justfile targetvllm.patchline numbers for 0.23.0Why remove CUDA 13?
cu129 nvcc already supports sm100/sm120 (Blackwell). The cu13 build path was unnecessary and caused failures: cu13 apt packages hijack
/etc/alternatives/cudafromcuda-12.9tocuda-13.0(which has no nvcc), breaking TE CMake withNo CMAKE_CUDA_COMPILER could be found.Test plan
vllm.__version__ == '0.23.0'vllm.patchapply verified viapatch --dry-runon 0.23.0 base🤖 Generated with Claude Code