[DNM][CI bisect] Revert #58197 to identify FT E2E detection hang

[vllm-agent]

Bisecting the main-branch Fault Tolerance E2E hang (rank 0 never reports UNHEALTHY after injected fault; broke between daily #90513 and #90560).

This PR reverts #58197 (disable JIT warmup in eager mode — the FT lane runs --enforce-eager). Do not merge — bisect probe only; will be closed once the culprit is identified. Sibling probe: #58317 (revert #55891).

[khluu]

/ci run

[github-actions]

✅ Triggered Buildkite CI #90636 for commit d543e065c0cf.

[khluu]

/ci run

[github-actions]

✅ Triggered Buildkite CI #90788 for commit 9debc2f83492.

[khluu]

/ci run

[github-actions]

❌ This PR is 2 commits behind upstream main. Your branch must contain every commit currently on upstream main. No new CI build was started. Merge or rebase onto the latest main, then rerun /ci run. To test this branch at your own risk, use /ci run --allow-stale.

[khluu]

/ci run

[github-actions]

✅ Triggered Buildkite CI #90793 for commit c7b1e3a84384.

[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.

[yewentao256]

CC @mgoin

[vllm-agent]

Context for a forward fix (per @khluu's preference over this revert):

Empirically proven: reverting #58197 makes the FT lane pass (build #90793, 2/2 green); main fails deterministically since #58197 merged. So the trigger is specifically kernel_config.enable_jit_warmup=False under --enforce-eager.

What enable_jit_warmup=False gates (from code read): the main JIT warmup batch (vllm/model_executor/warmup/kernel_warmup.py:170), plus MLA (mla_attention.py:657) and sparse-indexer (sparse_attn_indexer.py:908) warmup paths.

Observed failure shape: in test_injected_fault_retry_recovers_all_ranks, the faulted engine (DP1) goes UNHEALTHY correctly; the survivor (DP0) shows NIXL-EP dispatch timeouts at ~30s but its busy-loop DP sync never raises, so it never reports UNHEALTHY within the deadline. The lane runs with --enforce-eager + 2×DP + nixl_ep.

Not yet pinned: which specific warmed kernel the survivor's fault-time path depends on (the failing log shows router-GEMM/FlashInfer warmups still ran — so it's specifically the gated batch that matters). A forward fix could be as narrow as: don't disable JIT warmup when parallel_config.enable_fault_tolerance is set — but worth confirming that's sufficient via the FT lane before landing.

[njhill]

I think we can just drop --enforce-eager from that test, I will open a PR for that

[njhill]

I opened #58593 which should hopefully forward-fix the test, implementing the bot's suggestion.

A forward fix could be as narrow as: don't disable JIT warmup when parallel_config.enable_fault_tolerance is set — but worth confirming that's sufficient via the FT lane before landing.