[Thangnguyenvn98]

Overview

Multimodal Models (Standard) 4 runs as one long job because one of its commands, Whisper, starts with cd .. &&, and automatic sharding only plans steps whose test commands are plain pytest. This PR makes that command plain pytest and turns on automatic_shard for the step and its AMD mirror. It replaces the hand split in #57065.

Claims

• The Whisper command runs from the step's working directory, like the other three:
  pytest -v -s models/multimodal/generation/test_whisper.py -m core_model
• The cd .. and VLLM_WORKER_MULTIPROC_METHOD=spawn are no longer needed:
  • cd .. dates from [CI/Build] Reorganize models tests #17459, when a spawned test process could only import tests.* from the repo root. spawn_new_process_for_each_test now puts the repo root on PYTHONPATH itself ([CI] Add tests for cudagraph #27391), and test_common.py already runs spawned tests from tests/ in Standard 1–3.
  • test_whisper.py sets VLLM_WORKER_MULTIPROC_METHOD=spawn in an autouse fixture ([CI] Whisper logprobs tests #30504).
• The step and its mirror each run as parallel jobs planned from main's per-file timings. Every test runs exactly once across the shards:

Step	Planned shards
multi-modal-models-standard-4-other-whisper	3 (test_vit_cudagraph.py split by test)
amd-multi-modal-models-standard-4-other-whisper (mirror)	3

• If planning fails, the step runs as one job, as it does today.

Validation

• Planned both on this branch with ci-infra main's runtime_shard.plan_step, using main's timings on each step's own queue.
• Rendered both with the pipeline generator and ran every shard's commands against a mock test tree, as a PR build and with main's pytest -p ci_otel wrapper. Every test ran in exactly one shard, with the same env.
• /ci run checks the real question: whether Whisper passes without cd .. on both lanes.

Details

• Decision to check: dropping cd .. instead of moving Whisper to its own step. If the build shows the old import failure, the fallback is a separate Whisper step with working_dir: /vllm-workspace.
• Kept on purpose: the step key, even though "whisper" stays in it, because the planner looks up timings by key. Per-job timeouts are unchanged.

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

• AI assistance was used during the creation of this PR.

• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

• Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

• Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.

🤖 Generated with Claude Code

[Thangnguyenvn98]

/ci run

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[github-actions]

✅ Triggered Buildkite CI #96011 for commit 74e8121b8687.