[nvpohanh]

[by Claude Code]

Motivation

The cookbook's Agentic Long-Context Deployment sections (#42726, #43083, #43799, #43983, #44065, #44228, #44654, #44712, #45039, #45111, #45311, #45513, #45749, #46005, #46506, #46547, #46807, #46948, #47027, #47214, #47464) mirror SemiAnalysis InferenceX AgentX submissions as raw launch commands, each linked to the run that measured it. The configurations are mirrored from SemiAnalysis submissions. This PR:

1. Reverts [Doc] cookbook: Add Qwen3.8-Flash-Next B300 NVFP4 AgentX recipes #47560 and overlays its source PR instead. [Doc] cookbook: Add Qwen3.8-Flash-Next B300 NVFP4 AgentX recipes #47560 added a hand-written module (qwen3.8-flash-next-b300.jsx) and extra section text to the Qwen3.8-Flash-Next page for the configurations of InferenceX #4683. This PR reverts it and overlays InferenceX #4683 through the sync. The same eight B300 configurations are now generated, pinned to the run [Doc] cookbook: Add Qwen3.8-Flash-Next B300 NVFP4 AgentX recipes #47560 cited:

   • TP1 at c1, 2, 12, 24, 48, 80.
   • TP4·EP4 at c1, 2.

   They replace the B300 key's earlier RadixArk-checkpoint configurations, as refactor: Veteran-level full-stack improvements to code quality #4683 does in nvidia-master.yaml.

2. Syncs to InferenceX main (c5e8b9b3). These merged lanes now come from their dashboard runs:

   • MiniMax-M3.1-Flash GB300 SGLang: InferenceX #4676, run 41054395691, 2026-10-31.
   • MiniMax-M3.1-Flash GB300 Dynamo+SGLang agg and disagg: InferenceX #4698, run 41271829525, 2026-11-02. It follows InferenceX #4686.
   • DeepSeek-V4.1-Flash GB300 Dynamo+SGLang: InferenceX #4595 plus the append-only points of InferenceX #4679, run 41081085818, 2026-10-30.

   [Doc] cookbook: Update MiniMax-M3.1-Flash AgentX launches #47623, [Doc] cookbook: Add MiniMax-M3.1-Flash GB300 PD AgentX points #47699 and [Doc] cookbook: Sync MiniMax-M3.1-Flash GB300 AgentX points #47791 had edited the MiniMax data by hand for InferenceX #4675 / #4676 / #4686 / #4698. All four have since merged, and the sync regenerates the file from InferenceX main and the dashboard.

3. Overlays seven more InferenceX PRs:

   • Muse Spark Open, retuned per concurrency: B300 on the 20261030 nightly, GB300, B200 and GB200.
   • MiniMax-M3.1-Flash: GB200 Dynamo+SGLang config, and new B300 TP2 / TP4 points.
   • DeepSeek-V4.1-Flash GB300 SGLang on the 20261030 nightly.

Source

InferenceX main (c5e8b9b3), with these InferenceX PRs overlaid:

PR	Page	Lanes	Pinned to
#4726 MiniMax-M3.1-Flash B300 SGLang: TP2 and TP4 points	MiniMax-M3.1	minimaxm31flash-fp4-b300-sglang-agentic-dspark	no /use yet: in progress at head 0979dfd8
#4725 Muse Spark Open B300 SGLang, nightly-dev-cu13-20261030-302b78cc, retuned per concurrency	Muse Spark Open	musespark-fp4-b300-sglang-agentic-dflash	run 41328422543, named by the PR's /use comment (head 7516df00, unchanged), verified
#4724 Muse Spark Open GB300 SGLang, retuned per concurrency	Muse Spark Open	musespark-fp4-gb300-sglang-agentic-dflash	run 41327192200, named by the PR's /use comment (head 98ea294c, unchanged), verified
#4707 Muse Spark Open B200 SGLang, retuned per concurrency	Muse Spark Open	musespark-fp4-b200-sglang-agentic-dflash	run 41297696082, named by the PR's /use comment (head c965ca2e, unchanged), verified
#4695 MiniMax-M3.1-Flash GB200 Dynamo+SGLang config	MiniMax-M3.1	minimaxm31flash-fp4-gb200-dynamo-sglang-agentic-dspark-agg, -disagg	run 41271615328, named by the PR's /use comment (head fa292b90, unchanged), verified
#4683 Qwen3.8-Flash-Next B300 SGLang, NVIDIA NVFP4 checkpoint	Qwen3.8-Flash-Next	qwen3.8next-fp4-b300-sglang-agentic-mtp	run 41117066222, named by the PR's /use comment at d06564bc, verified. The PR head has moved to 6f2f7852 since; the cookbook shows the measured commit.
#4670 Muse Spark Open GB200 SGLang, retuned per concurrency	Muse Spark Open	musespark-fp4-gb200-sglang-agentic-dflash	run 41096285400, named by the PR's /use comment (head c9eb455f, unchanged), verified
#4603 DeepSeek-V4.1-Flash GB300 SGLang, 20261030 nightly, low-latency settings, TP4 c1 megakernel	DeepSeek-V4.1	dsv41flash-fp4-gb300-sglang-agentic-dspark	run 41261902147, named by the PR's /use comment (head 41d29b28, unchanged), verified

"In progress" lanes are rendered at the PR head and labeled "In progress in SA InferenceX AgentX". A later sync pins them once the PR gets a /use run or merges.

• MiniMax-M3.1-Flash B300 is all in progress. InferenceX #4726 changes the B300 key, which takes all of its points. That includes the points main measured in InferenceX #4675's run on 2026-10-31. They show as verified again once Use underlying device dynamically in MiniCPMV2_6 and Mllama models #4726 gets a /use run or merges.
• No longer overlaid:
  • Muse Spark Open GB300 musespark-fp4-gb300-sglang-agentic-dflash-multinode and -dflash-nightly1019 (InferenceX #4454, still open) leave the page.
  • InferenceX #4333 (Qwen3.5 GB300 Dynamo+SGLang) was closed without merging. Its retirement of qwen3.5-fp4-gb300-dynamo-sglang-agentic-agg-pareto and -pp-pareto no longer applies, so both return from their measured main runs. -agg and -disagg are back on their main runs.
  • InferenceX #4595 merged and comes from main (see above).
• DeepSeek-V4 page unchanged, as in [Doc] cookbook: Refresh SemiAnalysis AgentX launches to InferenceX main ea8b3439 #47464: InferenceX #4664 retired DeepSeek-V4-Pro.
• Not mirrored (unmeasured, unchanged): musespark-fp4-h100-sglang-agentic-dflash and qwen3.5-fp8-h100-sglang-agentic. The dashboard has no rows for them.
• InferenceX main was frozen for this PR at c5e8b9b3. Later merges are left for the next refresh.

What this PR does

Page	Lanes	Deployments	Points	Hardware
MiniMax-M3.1	8 (was 6)	36 (was 30)	118 (was 97)	GB300, GB200, B300, B200
GLM-5.2	14	34	75	GB300, GB200, B300, B200, H200
GLM-5.3	1	1	10	B300
DeepSeek-V4	unchanged
DeepSeek-V4.1	10	23 (was 21)	67 (was 65)	GB300, GB200, B300, B200, H200, H100
Qwen3.5	18 (was 16)	52 (was 51)	222 (was 220)	GB300, GB200, B300, B200, H200, H100
Qwen3.8-Flash-Next	4	6 (was 5, plus #47560's hand-written module)	27 (was 24)	GB300, B300, B200, H200
Muse Spark Open	6 (was 8)	9 (was 11)	56 (was 70)	GB300, GB200, B300, B200, H200

• Three new mirrored settings (kept as submitted):
  • DYN_TCP_CONNECT_TIMEOUT: the Dynamo request-plane TCP connect timeout, on the MiniMax-M3.1-Flash GB200 / GB300 Dynamo+SGLang lanes.
  • --prefill-decode-interval-min-bs: applies --prefill-decode-interval only at or above that running batch size ([Feat] scheduler: Gate --prefill-decode-interval on running batch size #44001), on Muse Spark Open GB300.
  • --strip-thinking-cache: keeps only the prompt prefix of finished reasoning requests in the radix cache (Opt-in strip of thinking tokens from radix cache #23315), on Qwen3.8-Flash-Next B300 c80 (InferenceX #4683).
• Variant sweeps. The MiniMax-M3.1-Flash B300 key measures up to five variants of one shape at the same concurrency, for example TP2 at c48. When no single worker flag tells them apart, each deployment is now labeled with the flags that do, e.g. --prefill-decode-interval 24 · --mem-fraction-static 0.93. Deployments that one flag already separated keep their labels.
• MiniMax GB300 Dynamo image. The MiniMax-M3.1-Flash GB200 / GB300 Dynamo+SGLang lanes use lmsysorg/sglang:nightly-cu134-20261030-e371881. It is an arm64-only image whose config reports SGLANG_BUILD_COMMIT=unknown. The tag suffix and the installed sglang-0.0.0.dev1+ge3718813e both identify SGLang e3718813, which is now used for its KV-tier checks.
• The GLM-5.2 and GLM-5.3 data change only in their shared source block (the InferenceX main SHA and the overlay list).

Validation

• Browser checks: every page this PR changes was rendered in a headless browser (mint dev, Mintlify as in CI).

  • Every hardware / checkpoint / deployment / concurrency / KV tier / router combination renders exactly the decoded commands: 3,116 command sets, 0 mismatches, 0 unseen, no browser errors.
  • Every tab was clicked, and all 161 "Copy link" URLs were reopened.
  • The section links from the Deployment panels and the copy buttons were also checked.

• Decode and KV-flag checks:

  • The component's own decoder reproduces every generated plan (3,116 router × KV-tier variants).
  • Every KV flag exists in its image's SGLang commit (226 image × flag pairs).

• Arg-parse: all 574 commands not validated in an earlier round parse:

  Image	Commands	Parsed with
  dev-minimax-m31-flash	176	the image
  nightly-dev-cu13-20261030-302b78cc (tag and digest)	225	the image, and again with SGLang 52604e1d (its arm64 build's commit) for the GB300 lanes
  nightly-cu134-20261030-e371881 (arm64 only)	97	SGLang e3718813 source and the image's bundled ai-dynamo 1.7.0.dev20261030, on the amd64 20261030 nightly
  nightly-dev-cu13-20261029-c10e0e17	12	the image
  dev-muse-spark-open	12	the image
  nightly-dev-20261026-8252e3af	6	the image
  nightly-dev-cu13-20261028-586355a8	46	the image

• node scripts/check_cookbook_configs.mjs, mint broken-links --check-anchors --check-redirects, mint validate and pre-commit run --all-files (in lmsysorg/sglang:dev-cu13) pass.

---

CI States

Latest PR Test (Base): ✅ Run #41464298005
Latest PR Test (Extra): ❌ Run #41464297465
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

🤖 Generated with Claude Code

[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 approved

@nvpohanh approved a Mintlify preview for commit 3a3825d.

[mintlify]

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

Project	Status	Preview	Updated
sglang-doc	🟢 Ready	View Preview	Nov 2, 2026, 3:06 PM

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

[nvpohanh]

Preview looks good. merge to unblock SA submissions