[diffusion] fix --warmup silently downgrading server-based warmup to request mode

[mickqian]

Motivation

sglang serve --warmup is meant to use server-based warmup — the serve entrypoint injects a default warmup_mode="server" "for production". But it silently downgrades to request-based warmup, and combined with --warmup-resolutions it disables warmup entirely. The --warmup flag can thus effectively turn warmup off, leaving the first real request to pay the full eager cold-start (observed as ~2× latency in diffusion benchmarking).

Root cause

serve injects warmup_mode="server" via default_args (runtime/entrypoints/cli/serve.py), but default_args are not recorded in _explicit_arg_names. So in _adjust_warmup (runtime/server_args.py), when --warmup is explicit (legacy_explicit=True) but --warmup-mode is not (mode_explicit=False), the guard if mode_explicit or not legacy_explicit: evaluates to False — the injected warmup_mode="server" is never applied, and server_warmup falls back to its dataclass default False (i.e. request mode).

--warmup --warmup-resolutions WxH then becomes a dead zone:

• synthetic server warmup requires server_warmup=True → never runs;
• request-based warmup bails out when warmup_resolutions is not None → never runs.

→ no warmup runs at all.

Fix

In _adjust_warmup, when a legacy --warmup/--server-warmup enabled warmup without an explicit --warmup-mode, honor the resolved warmup_mode (e.g. serve's "server" default) for the server-vs-request split, instead of silently staying request-based. --warmup false keeps warmup off and is untouched.

Test

Adds unit tests in TestWarmupModeNormalization:

• legacy --warmup with mode defaulted to server → resolves to server_warmup=True;
• the --warmup --warmup-resolutions dead-zone regression → server-based warmup runs.

Existing warmup tests (--warmup false, explicit --warmup-mode, disagg, bare serve) still pass.

Verification (H100)

With this fix, serve --warmup runs server-based warmup, and a diffusion benchmark's measured latencies match the prior client-side-warmup baseline instead of the ~2× cold-start regression seen when warmup silently didn't run — e.g. Z-Image-Turbo 0.74s vs 0.65s baseline (was 13.46s with no warmup), FLUX.1-dev 4.80s vs 4.68s, Ideogram-4 5.26s vs 5.19s.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ✅ Run #28293800395
Latest PR Test (Extra): ❌ Run #28293800347

[gemini-code-assist]

Code Review

This pull request updates the server argument resolution logic to correctly handle legacy --warmup and --server-warmup flags when no explicit --warmup-mode is provided. It ensures that the resolved warmup_mode (such as the default 'server' mode) is honored instead of silently downgrading to request-based warmup. Additionally, unit tests have been added to verify this behavior and prevent regressions. There are no review comments, and I have no additional feedback to provide.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[mickqian]

/tag-and-rerun-ci