[Duyi-Wang]

Motivation

The MI355X DeepSeek-V4-Pro-0813 PD cookbook still documents c1 with two request slots and decode graph sizes 1–2. Its Docker commands also select an older image despite enabling mxscale GEMM and power-of-two routing settings.

This completes the cookbook follow-up to #44454 identified in InferenceX #4186. Source configuration: recipe at 6ad6674d9de2.

Modifications

• Document eight request slots on both c1 workers and decode graph sizes 1–8, while keeping client concurrency at one. Explain the exception to the strategy presets' 2 × concurrency sizing rule.
• Select lmsysorg/sglang-rocm:v0.5.22-rocm720-mi35x-20261019 for both Pro and Pro Official FP4 PD prefill, decode and router Docker commands, and update the matching example. Both variants share one image; single-node images remain unchanged.
• Remove submission/PR-following descriptions from the PD documentation and config comments; keep provenance in this PR description.

This touches §3.7 and its deployment guidance/config resolver. #45268 addresses the separate single-node §3.6 recipe.

Validation

• git diff --check and pre-commit on both changed files: passed.
• node docs/scripts/check_cookbook_configs.mjs: passed.
• node --test docs/scripts/tests/*.test.mjs: 38/38 passed.
• Evaluated all 18 combinations of Pro/Pro Official × three strategies × three PD roles: both variants select the new image, and every other resolved field matches the base revision.
• Compared c1 admission limits, decode graph sizes and TP sizing with the pinned source recipe: passed.
• mint broken-links --check-anchors --check-redirects (CI-pinned mint 4.2.559): passed.
• mint validate: blocked by the pre-existing extra }, at docs/src/snippets/configs/popular-models.jsx:100, which prevents its imports in the two landing pages. The unchanged file at the base revision also fails JavaScript parsing.
• Verified the public Docker Hub tag is active, digest sha256:8cb29fc407c2e46bdb1866c5edb91f109ac1acbbcb6905089beaa6c79b0813a2.

Accuracy Tests

No new GPU runs; documentation and command-generation image selection only. The original Pro EAGLE/MTP variant has not been revalidated on the newer image; the linked run covers Pro Official DSpark. Existing source performance/eval evidence: InferenceX run 39244119974.

Speed Tests and Profiling

Not rerun. No runtime implementation changes.

Checklist

• Format code with pre-commit.
• Validate affected command generation and run existing documentation tests.
• Update documentation.
• Follow SGLang code style.

---

CI States

Latest PR Test (Base): ✅ Run #39363751285
Latest PR Test (Extra): ❌ Run #39363751433
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

[1am9trash]

#45268 addresses the separate single-node §3.6 recipe.

Just a clarification, 3.6 section is included in #45070.

[1am9trash]

LGTM