Add vLLM x TorchAO integration workflow

[huydhn]

This adds a workflow to run vLLM x TorchAO tests with the latest vLLM main. The setup is vLLM main x PyTorch stable x FBGEMM stable x TorchAO PR/main commits.

If there are not enough H100 capacity and you observe queueing, please let me know. We might need to tweak the workflow to run on H100 only when pushing to main.

Testing

https://github.com/pytorch/ao/actions/runs/16535494171/job/46769051375

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/2610

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit a6b2aaa with merge base 5fe4ebd ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[huydhn]

@jerryzh168 I run the same logic to install fbgemm-gen-ai from nightly and build TorchAO, but running pytest currently ends up with this error loading fbgemm https://github.com/pytorch/ao/actions/runs/16535051236/job/46767869902?pr=2610#step:7:400. Do you know what I did wrong here?

I need to install fbgemm-gen-ai from stable to avoid rebuilding it. I think we could stay with that, but let me know if it's ok. In this setup we have, vLLM main x PyTorch stable x FBGEMM stable x TorchAO PR/main commits

[huydhn]

@pytorchbot drci

[jerryzh168]

current torchao main doesn't work with fbgemm gpu genai stable I think, maybe just use stable for everything and we'll make sure next stable torchao release can work with fbgemm gpu genai stable

[huydhn]

current torchao main doesn't work with fbgemm gpu genai stable I think, maybe just use stable for everything and we'll make sure next stable torchao release can work with fbgemm gpu genai stable

Interesting, the tests pass https://github.com/pytorch/ao/actions/runs/16537610493?pr=2610, it means that torchao main is working with vLLM main. Is that what we are looking for here? I guess if we need to test torchao vs fbgemm gpu genai, it makes more sense to have another workflow to test only these twos?

[jerryzh168]

current torchao main doesn't work with fbgemm gpu genai stable I think, maybe just use stable for everything and we'll make sure next stable torchao release can work with fbgemm gpu genai stable

Interesting, the tests pass pytorch/ao/actions/runs/16537610493?pr=2610, it means that torchao main is working with vLLM main. Is that what we are looking for here? I guess if we need to test torchao vs fbgemm gpu genai, it makes more sense to have another workflow to test only these twos?

OK it could be because we haven't landed everything we want in main yet, and some of our new changes would require fbgemm main

we do have tests to test torchao + fbgemm gpu genai (nightly) in our CI

[jerryzh168]

nit: nightly might be enough I think

[jerryzh168]

fbgemm gpu genai nightly would require pytorch nightly I think, and with that we probably need vllm to depend no torch nightly, is that doable with docker? or docker is build with stable pytorch?