[Thangnguyenvn98]

Overview

Enrolls four Nightly steps and three of their AMD mirrors in automatic_shard, continuing the earlier enrollment batches (#61999, #62039, #64115). This PR only sets the flag; it changes no tests or commands. A step that fits in one job runs as one job, unchanged, and splits without another CI change once its tests outgrow it.

Claims

Step	Planned shards	AMD mirror
multi-modal-models-extended-pooling	3	2
lm-eval-deepseek-v32-tp-8xh200	1	1
lm-eval-deepseek-r1-tp-8xh200	1	1
lm-eval-hisparse-nightly-4xb200	1	none

• Each job runs the same tests, environment and timeout as on main. One gsm8k config takes about as long as a whole shard, so the planner keeps the lm-eval steps whole for now.

Validation

• Planned each step and mirror with ci-infra main's runtime_shard.plan_step against main's timings, and rendered each with the pipeline generator as a PR build and in main's OTel mode.
• /ci run (98071) on this head, plus a build scoped to the three AMD mirrors it leaves behind the block step (98079): every job passed. Extended Pooling ran its planned 3 shards (mirror: 2) with the same pass and skip counts as main; the lm-eval steps ran as one job each, as on main.

Details

• Extended Pooling is back. An earlier build of this PR failed it on the ColUltraVBERT checkpoint; main fixed that in [CI][Bugfix] Pin the ColUltraVBERT test checkpoint to the revision vLLM supports #64189 and its nightly passes the step again.
• Cost. Extended Pooling's machine time rises by about a third: each extra shard repeats job setup and the model loads its test files share. In exchange its wall time halves. It runs nightly, so we keep the shorter wall time.
• Dropped from this batch. lm-eval-humming-b200-act fails on main from a regression there. It comes back in a later batch once main passes it.
• Timeouts unchanged (human). timeout_in_minutes stays per job. Fallback-aware timeouts belong in a central timeout setting.
• AI assistance. Claude Code prepared the change and this description; the submitter reviews every changed line.

🤖 Generated with Claude Code

---

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

[github-actions]

✅ Triggered Buildkite CI #97804 for commit 4756f6bab452.

[Thangnguyenvn98]

/ci run

[github-actions]

✅ Triggered Buildkite CI #98071 for commit 55b6a8728932.

[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.

[Thangnguyenvn98]

/ci run

[github-actions]

✅ Triggered Buildkite CI #98767 for commit 6dc1ef0f957f.