[Bugfix][CI/Build] Fix CUDA 11.8 Build

[LucasWilkinson]

Dont build 9.0a for scaled_mm_c2x since it's outside of cuda 12.0 guard and won't help perf that much that anyways.

The issue here was that for CUDA 11.8 if we were building for 9.0 we wouldn't build scaled_mm_c3x so we would instead try to build scaled_mm_c2x for all versions, i.e. "7.5;8.0;8.6;8.9;9.0;9.0a", this is incorrect though since 9.0a isn't supported by 11.8. We can just drop 9.0a for scaled_mm_c2x since scaled_mm_c2x won't take advantage of the 9.0a features anyways.

(and fix c3x error message false reporting that there were no compatible arches when on 11.8)

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[simon-mo]

Thanks @LucasWilkinson, is this ready for review?

[LucasWilkinson]

Thanks @LucasWilkinson, is this ready for review?

Basically just waiting for my docker build tests to finish to confirm the fix, they are slow haha

[LucasWilkinson]

Confirmed, this builds (i.e. this PR)

FROM pytorch/pytorch:2.4.0-cuda11.8-cudnn9-devel AS build

ARG torch_cuda_arch_list='7.0 7.5 8.0 8.6 8.9 9.0'
ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}

RUN apt update && apt install gcc g++ git -y && apt clean && rm -rf /var/lib/apt/lists/*

ENV PATH=/workspace-lib:/workspace-lib/bin:$PATH
ENV PYTHONUSERBASE=/workspace-lib

RUN pip install git+https://github.com/neuralmagic/vllm.git@5a7b00e7a6377ca7971de3ca762583a9153f4a55 --no-cache-dir --user -v

and this fails (i.e. main):

FROM pytorch/pytorch:2.4.0-cuda11.8-cudnn9-devel AS build

ARG torch_cuda_arch_list='7.0 7.5 8.0 8.6 8.9 9.0'
ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}

RUN apt update && apt install gcc g++ git -y && apt clean && rm -rf /var/lib/apt/lists/*

ENV PATH=/workspace-lib:/workspace-lib/bin:$PATH
ENV PYTHONUSERBASE=/workspace-lib

RUN pip install git+https://github.com/vllm-project/vllm.git@e9d517f27673ec8736c026f2311d3c250d5f9061 --no-cache-dir --user -v

[tlrmchlsmth]

LGTM, thanks for the fix!

[mgoin]

Thanks for the quick fix!