[Feature | CI] Added a github action to build wheels

[Danielkinz]

Changes

Created a new action called "Create Release" (publish.yml) which automatically detects a branch of the format "release/<version>", creates a release with same version, build and uploads wheels for said release.

Supported Versions:

The following covers which environments the wheels cover:

• Linux with glibc 2.31 or newer (ex. ubuntu 20.04 or newer) - github doesn't have older runners
• Cuda toolkit 11.8 - github runner failed to install torch for cuda 11.7. Older versions not supported by torch >=2.0.0.
• Python 3.8, 3.9, 3.10, 3.11 - torch doesn't support older versions

Notes

• The github action can be edited to fit the main repository's release pattern.
• Example runs and results can be seen in the forked repository

Next Steps

• Replace manual build call with cibuildwheel - could possibly add more flexibility to the wheel versions.
• Add support for windows/macos if needed.
• Add cuda 12.x once torch 2.1.0 becomes stable

[WoosukKwon]

Wonderful! Thank you so much for this amazing work @Danielkinz !
I've checked that the wheels are automatically generated by the GitHub action, and the wheels can be installed very smoothly on python 3.8, 3.9, 3.10, and 3.11. It also works on the NVIDIA PyTorch containers with CUDA 12 installed. 👍

I only request small changes:

1. Could you change the names of the generated wheels from, e.g., vllm-0.1.3-cp310-cp310-linux_x86_64.whl to vllm-0.1.3-cp310-cp310-manylinux1_x86_64.whl which is compatible with the naming convention in PyPI?
2. When creating a release, could you also include the auto-generated release note?
3. Could you change the GitHub action trigger to any new v* tag?

After all, this is very nice work. Thank you so much again @Danielkinz

[Danielkinz]

Thank you for the review @WoosukKwon :)

About the requested changes:

1. Changing the name should pose no issue at all. I already experimented with it a bit. Also, I believe I saw an action to automatically upload to PyPI. Would you like me to check it out as well?
2. How does the autogeneration work currently? I didn't see a .github/release.yml file.
3. Switching to trigger on tag should be simple. If you prefer its also possible to trigger on release to fill in the wheels, which might solve 2 as well

[WoosukKwon]

Hi @Danielkinz,

1. Yeah, it would be awesome if the GitHub action can also rename and upload the wheels to PyPI.
2. I think it is possible by using GitHub CLI. Could you refer to this link?
3. Yes, please switch the trigger. I personally feel the v* tag trigger is a bit more intuitive.

In summary, the ideal workflow for us is that the GitHub action is triggered when a new v* tag is created, then it builds python wheels, creates new release with the wheels and an auto-generated release note, and uploads the wheels to PyPI. However, we can skip 2 (release notes) for now if it is not easy.

[Danielkinz]

Alright. I'll take a look at it over the next couple of days :)

[WoosukKwon]

Hi @Danielkinz, what's going on this PR? Please let us know if you need any help.

[Danielkinz]

Ah sorry about that. Last night I updated the fork and moved some branches around and the fork's main branch ended up on the same commit as the main repo's. Seems like github automatically closes PR's when that happens.
The requested features are almost done and are currently in testing. The only feature I'm having a bit of trouble with is the PyPI auto upload since its more difficult to test it. (since it requires to set up a test.pypi project)

[WoosukKwon]

Hi @Danielkinz, thanks for sharing the status! Then, I think we can manually upload the wheels for now, and implement the feature later. WDYT? This PR is already super awesome.

[Danielkinz]

As long as you're fine with it. Theoretically the PyPI auto upload should work, I simply can't test it easily. If you'd like you can try it out on the main branch with test.pypi next time there's a release on the main branch, it shouldn't affect the rest of the pipeline.
I currently encountered another problem - it seems like nvidia has updated their apt repo last night, which causes an error when trying to install their libraries:
"E: Failed to fetch https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/Packages.gz File has unexpected size (1127072 != 1126689). Mirror sync in progress? [IP: 152.195.19.142 443]"
Their repo: https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/
For now I can't continue testing the features since this error kills the pipeline. It should be a temporary problem.

[WoosukKwon]

As long as you're fine with it. Theoretically the PyPI auto upload should work, I simply can't test it easily. If you'd like you can try it out on the main branch with test.pypi next time there's a release on the main branch, it shouldn't affect the rest of the pipeline.

Yes, let's have it in a future PR, as many users are waiting for this PR to be merged. 😃 Could you just let me know if 1) the release notes are now auto-generated, and 2) the wheels are renamed to the format like vllm-0.1.3-cp310-cp310-manylinux1_x86_64.whl? I believe these two are also optional and should not block this PR though.

[Danielkinz]

Yes, both are implemented. After I test them the PR should be good to go.

[WoosukKwon]

@Danielkinz Great, please let me know if the PR is ready.

[Danielkinz]

@WoosukKwon The PR is ready :)
Seems like the issue with nvidia's repo was resolved last night so I could test it out.

[WoosukKwon]

@Danielkinz LGTM. Thanks for the super awesome work! This will save many users from the tedious installation issues. Great work!