[CI/Build] Dockerfile.cpu improvements

[dtrifiro]

• use requirements-build.txt to install build requirements
• use cache bind mounts to speed up builds
• use ccache
• build wheel and install it instead of using setup.py install

[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 consists a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of default ones by unblocking the steps in your fast-check build on Buildkite UI.

Once the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge).

To run full CI, you can do one of these:

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

🚀

[dtrifiro]

/ready

[dtrifiro]

A significant issue here is that the torch version installed by requirements-cpu.txt will override whatever is defined in requirements-cpu.txt.

[mgoin]

@bigPYJ1151 @zhouyuan does this look okay by you?

[bigPYJ1151]

Overall LGTM, thanks for the improvement!

The only thing I concerned is using requirements-build.txt will install torch with many CUDA dependencies, which looks a bit redundant and increases download time.

[dtrifiro]

It will not: PIP_EXTRA_INDEX_URL is set before the RUN command, resulting in the following dependencies being installed:

Package           Version
----------------- ---------
cmake             3.30.2
filelock          3.15.4
fsspec            2024.6.1
Jinja2            3.1.4
MarkupSafe        2.1.5
mpmath            1.3.0
networkx          3.3
ninja             1.11.1.1
packaging         24.1
pip               24.1.2
setuptools        70.3.0
sympy             1.13.1
torch             2.4.0+cpu
typing_extensions 4.12.2
wheel             0.43.0

One possible issue might come up with some package managers such as uv:

export UV_EXTRA_INDEX_URL=https://download.pytorch.org/whl/cpu
uv pip install -r requirements-build.txt

will fail due to the way uv resolves local version identifiers: in order for this to succeed, the +cpu specifier should be explicitly added (torch==2.40+cpu).

[dtrifiro]

Note that this PIP_EXTRA_INDEX url also works for PEP517 style builds, so running

export PIP_EXTRA_INDEX_URL=https://download.pytorch.org/whl/cpu 
export VLLM_TARGET_DEVICE=cpu 
pip install git+https://github.com/vllm-project/vllm

will also install the cpu version when creating the isolated build venv.

[maxdebayser]

Overall the changes look good to me. There's only one detail I asked for more clarification about.

[maxdebayser]

Do the files in dist stick around in the container image? Can they be removed with && rm -r dist or is there some dependency on these files after pip install?

[dtrifiro]

Yeah, that's correct, good catch. I did not bother too much about cleaning the image as I think this is clear this is mostly a CI image.

If there's an interest in making this production ready, we can set up multi-stage builds and have a final deployment stage without dev/build dependencies.

I can remove this if others think this is worth removing in this PR

[mgoin]

This isn't a big overhead, we can remove later