Embed git commit hash into Python source

[dbarbuzzi]

This PR performs a simple embed of the current git commit hash into the Python source:

1. The hash itself is defined in the __githash__ __commit__ var in vllm/version.py alongside __version__
2. vllm/__init__.py is updated to (a) import __githash__ __commit__ and (b) include it in the __all__ list

As a result, you can now do the following:

import vllm
print(vllm.__version__)  # already present
print(vllm.__commit__)  # now available

Test Plan
New test added which minimally verifies that this new variable is defined to a value with a bare-minimum length.

[derekk-nm]

Thank you! Thank you! Thank you!

[robertgshaw2-neuralmagic]

This looks good.

Can you add a test that this is working properly?

[mgoin]

Could we register it as vllm.__commit__? I believe this is what Simon requested upstream

[dbarbuzzi]

Could we register it as vllm.commit? I believe this is what Simon requested upstream

Done @mgoin

[dbarbuzzi]

Can you add a test that this is working properly?

@robertgshaw2-neuralmagic I've added a basic test here (checks that the variable is defined): https://github.com/neuralmagic/nm-vllm/blob/simple-githash-embed/tests/test_embedded_commit.py

Let me know if you want something more in-depth.

[dbarbuzzi]

One of the LM-EVALs failed (think it’s the same one that failed in a nightly, marlin non-determinism?):

FAILED tests/accuracy/test_lm_eval_correctness.py::test_lm_eval_correctness[TechxGenus/Meta-Llama-3-8B-Instruct-GPTQ] - assert False
 +  where False = <function isclose at 0x7ad222e75bf0>(0.688, 0.648, rtol=0.05)
 +    where <function isclose at 0x7ad222e75bf0> = numpy.isclose

[mgoin]

I don't like the fact that we can't rely on vllm.__commit__ existing in the general case since this only happen during build. Could you at least add an empty __commit__ = None to version.py so the attribute is checkable?

This is also a bit specific to our build process. I thought we would integrate this more into the vllm build process natively. Something like this would work towards that: https://github.com/thuml/depyf/blob/3574bc0f12d9baad751354b19c432c822f8d321f/setup.py#L7-L16

[dbarbuzzi]

@mgoin I’ve updated the PR to embed it during setup.py. All plumbing is added to the __init__.py and version.py files with a placeholder for the value in the latter file.

However, my understanding is that we want this in all builds (nightly, release, etc.), I diverged from that example slightly during setup.py and instead it still directly calls the function that does the embedding so that version.py is re-written with the placeholder replaced. The way the example has it, the commit would be part of the version/tag that is present in the metadata on e.g. PyPI/pip which, to my knowledge, is not the path we want (let me know if that’s incorrect).

[andy-neuma]

just some minor nits.

[mgoin]

Looks perfect to me, thanks!

[mgoin]

Nit: I do think this is a bit long as a placeholder or dummy value

[dbarbuzzi]

Any suggestions for an appropriately short one? Maybe COMMIT_PLACEHOLDER, or is that still long?