[Core] Disable JIT warmup in eager mode

[mgoin]

Purpose

Make --enforce-eager also disable JIT kernel warmup. Eager mode already opts out of torch compilation and CUDA graph capture, so paying the up-front JIT compilation cost is unexpected and particularly expensive for short-lived processes.

AI assistance was used to investigate the startup regression and prepare this change. The submitter reviewed the final diff.

Test Plan

/home/mgoin/.cache/pre-commit/repoba318kxi/py_env-python3.12/bin/ruff check vllm/config/vllm.py
git diff --check

No model evaluation is needed because this only changes eager-mode startup warmup behavior; model execution and outputs are unchanged.

Test Result

All checks passed!
git diff --check: passed

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[LucasWilkinson]

Thank you!

[mgoin]

/ci run

[github-actions]

✅ Triggered Buildkite CI #90496 for commit 209f4e2c95fa.

[gau-nernst]

Can we discuss this? I think JIT warmup should stay on by default under --enforce-eager. The two flags are orthogonal: e.g. in PD serving, P workers often run with --enforce-eager, but still need JIT warmup (e.g. for CuTeDSL kernels) to avoid compile latency spikes while serving.

Today an explicit --kernel-config.enable-jit-warmup=True is overridden under eager, so it's not possible to opt-in again. The workaround is -cc.mode=none -cc.cudagraph_mode=NONE instead of --enforce-eager, but that's verbose and easy to miss.

Since this is a silent latency regression for deployments, I'd rather default to safe serving behavior and let short-lived processes opt out with --kernel-config.enable_jit_warmup=False (tests/CI can set it). That would revert this change and drop the fault-tolerance special case.