Add a workflow to run vLLM unit tests on H100

[huydhn]

This is a temporary workflow to run some vLLM unit tests on H100 until we have that runner on vLLM. As this is a workflow on PyTorch side and our H100 capacity is limited, we could only run the test:

• Manually on a vLLM PR:
  • For metamates, go to bunnylol oss pytorch/pytorch-integration-testing, and add yourselves to the repo which will grant you write permission to run the workflow
  • Go to https://github.com/pytorch/pytorch-integration-testing/actions/workflows/vllm-ci-test.yml to run the workflow
  • Set the branch name to refs/pull/<YOUR_PR_NUMBER>/head
• or periodically on vLLM main commit every 4 hours
  • You can get the vLLM main commit from the Docker image name in the following format public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:<COMMIT>, for example, GH logs

More tests can be added into .github/scripts/run_vllm_tests.sh, I use pytest -v models/multimodal/generation/test_maverick.py as a simple example. All the features like test sharding or model caching are not available at the moment, but this is a start.

Testing

https://github.com/pytorch/pytorch-integration-testing/actions/runs/16535983273/job/46770512704?pr=55#step:8:214

[yangw-dev]

lgtm!