[CI/Build] Test torchrun with 8 cards

[22quinn]

Purpose

Increase torchrun / external launcher mode CI coverage to prevent issues like #27502. Reverting #27502 would fail this test.

Test Plan

TP_SIZE=2 DP_SIZE=4 ENABLE_EP=1 torchrun --nproc-per-node=8 examples/offline_inference/torchrun_dp_example.py

Test Result

Pass

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[mergify]

Documentation preview: https://vllm--27548.org.readthedocs.build/en/27548/

[gemini-code-assist]

Code Review

This pull request adds a new CI test for torchrun with 8 GPUs to improve coverage for distributed inference, which is a great addition. The changes in torchrun_dp_example.py to allow configuration via environment variables are necessary for this. I've found a code quality issue in the example script where new module-level variables shadow later assignments, which can be confusing. I've left a couple of suggestions to improve clarity and avoid potential future bugs.

[houseroad]

do we have the capacity to run this?

[22quinn]

Seems not https://github.com/vllm-project/ci-infra/blob/a64d73f396ba8629e8fb9c5e5b933a5f87d1edc2/buildkite/pipeline_generator/utils.py#L32
Possible to add 8-card capacity? @hl475

[hl475]

if we want to use 8-card capacity, maybe try with H100? example PR https://github.com/vllm-project/vllm/pull/27113/files