[ChangLiu0709]

Motivation

InferenceX #4542 moves the MI355X AgentX lane qwen3.8next-fp4-mi355x-sglang-agentic-mtp (added by #4680, documented here by #47520) onto the 20261031 image and serves concurrency 1, 2, 4 and 8 with the FlyDSL monokernel (--enable-amd-monokernel). InferenceX requires the production serving settings on a merged SGLang cookbook page before the PR merges. The page documents one command, the #4680 default path, which no longer matches c1–c8.

Modifications

docs/cookbook/autoregressive/Qwen/Qwen3.8-Flash-Next.mdx, MI355X subsection only:

• Image: lmsysorg/sglang-rocm:v0.5.23-rocm10-mi35x-20261031@sha256:3867dc07ba2530f914e1f210ed829f7a572a75286c8c362dbc22cbecf9b5a7e6, built from SGLang e9c18259c7, AITER 285579b6 and sgl-amd-mono b10f3b71 (the pin since [Chore] [AMD] mono-kernel: Bump sgl-amd-mono for Qwen3.8-Flash-Next #47483).
• New "Monokernel (concurrency 1–8)" bullet: what --enable-amd-monokernel runs, what it needs (NEXTN 3/1/4, the Quark checkpoint, no --enable-linear-replayssm-spec), that it is experimental and that SGLang refuses unsupported setups at startup.
• PLE offload and batch-limit bullets updated: c1–c8 keep the table on the GPU and set both limits to the concurrency (1, 2, 4, 8); c12 and c16 keep the offload and 2x (24, 32).
• Two commands: Concurrency 12–16 (the existing command, unchanged) and Concurrency 1–8 (monokernel) (concurrency 8 shown).
• Not changed: the Deploy panel, the generated AgentX snippets, section 2.1.

Source of truth

• InferenceX PR: [AMD][MI35X] Qwen3.8-Flash-Next FP4 SGLang agentic MTP: 20261031 image, monokernel at c1-c8 / [AMD][MI35X] Qwen3.8-Flash-Next：20261031 镜像，c1-c8 启用 monokernel SemiAnalysisAI/InferenceX#4542, head reviewed 54b9aadc9efbb0b79053c845dcb38779467d85d7.
• Recipe: inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/qwen3.8next/sglang/mi355x-fp4-mtp/agentic.yaml (base, override_c8, override_c16); image in inferencex-e2e/configs/amd-master.yaml.
• Sweep: Run Sweep 41248881067 (head bdf93a7f1): all six agentic points and the c16 eval succeeded. The later head only merges main.
• SGLang main base: 16dd97df61.

Coverage

Each command was compared with the recipe's merged args for its point (c8 and c16): 0 missing, 0 conflicting, 0 extra settings besides --host and --port; the recipe's SGLANG_AITER_HONOR_EXPLICIT_MEM_FRACTION and AITER_GDR_FLYDSL are in both. Telemetry flags, PYTHONNOUSERSITE, SGLANG_TIMEOUT_KEEP_ALIVE and the benchmark env stay excluded, as in #47520. ROCM_QUICK_REDUCE_QUANTIZATION=NONE is the default and TP1 runs no all-reduce.

Open items

• InferenceX Fix the incorrect args in benchmark_and_profiling.md #4542 currently conflicts with its base; the recipe is expected to stay as reviewed. If its settings change, this page needs the same change before this PR merges.
• GSM8K runs only at c16 in the sweep (default path). c1–c8 have no accuracy eval there because the monokernel is selected by a CLI flag, not an env var.

Validation

• git diff --check: pass.
• Docs-only; no code, so no accuracy or speed tests.

---

CI States

Latest PR Test (Base): ✅ Run #41522596628
Latest PR Test (Extra): ❌ Run #41522596254
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

Tip

Base CI only covers the core cases, and its workload may not exercise your change. Use /rerun-test to run the tests your change touches, or copy this prompt for an agent:

Find the registered tests this PR affects and run them with /rerun-test, following https://github.com/sgl-project/sglang/blob/main/.agents/skills/rerun-test/SKILL.md

[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.

[mintlify]

Preview approval required

This pull request comes from a fork, so Mintlify won't build a preview until someone with write access approves commit e004b36.

Approve preview

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated
sglang-doc	🟢 Ready	View Preview	Nov 2, 2026, 8:37 PM

💡 Tip: Enable Automations to automatically generate PRs for you.

[zijiexia]

@claude review

[claude]

Code review found no issues

No high-confidence issues detected in this change.