[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) 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. Removes launches whose results SemiAnalysis purged. The merged Klaud-Cold image bumps InferenceX #4203, #4529, #4515, #4324 and #4449 had issues, so their results were purged from the InferenceX dashboard. [Doc] cookbook: Refresh SemiAnalysis AgentX launches with DeepSeek-V4 B200 DSpark6 disagg and Muse Spark GB300 points #46807 and [Doc] cookbook: Refresh SemiAnalysis AgentX launches to InferenceX main 67de4b34 #46948 mistook the missing rows for a dashboard problem and pinned these lanes to the PRs' /use runs. Each lane now shows the run the dashboard shows again (table below).
2. Adds the in-progress InferenceX PRs in the Source table: MiniMax-M3.1-Flash FP8 draft KV on B200 / GB200 / GB300, the DeepSeek-V4.1 GB300 DEP16 disagg HiCache lane, and DeepSeek-V4 B200 disagg DSpark6 HiCache.
3. Syncs to InferenceX main (253c38bf). Muse Spark Open B300 (InferenceX #4417) is now on the dashboard and comes from main.

Lanes back on their dashboard runs

Lane	Page	#46948 showed (purged)	Now (dashboard run)
dsv4-fp4-b200-dynamo-sglang-agentic-agg	DeepSeek-V4	InferenceX #4203	InferenceX #3257 (2026-09-22)
dsv4-fp4-gb200-dynamo-sglang-agentic-agg	DeepSeek-V4	InferenceX #4529	InferenceX #3701 (2026-10-12)
dsv41flash-fp4-gb300-dynamo-sglang-agentic-dspark-disagg	DeepSeek-V4.1	InferenceX #4515	InferenceX #4434 (2026-10-26)
dsv41flash-fp4-h100-sglang-agentic-dspark-dpa	DeepSeek-V4.1	InferenceX #4324	InferenceX #3345 (2026-09-24)
glm5.2-fp4-gb300-dynamo-sglang-agentic-agg	GLM-5.2	InferenceX #4449	InferenceX #4300 (2026-10-20)

From now on, the sync never pins a merged lane to a /use run the dashboard does not show. It lists such merges for review instead.

Source

InferenceX main (253c38bf), with these InferenceX PRs overlaid:

PR	Page	Lanes	Pinned to
#4600 MiniMax-M3.1-Flash GB300 SGLang FP8 draft KV	MiniMax-M3.1	minimaxm31flash-fp4-gb300-sglang-agentic-dspark	no /use yet: in progress at head 3b96f338
#4599 MiniMax-M3.1-Flash GB200 SGLang FP8 draft KV	MiniMax-M3.1	minimaxm31flash-fp4-gb200-sglang-agentic-dspark (all but conc 1)	no /use yet: in progress at head 84ee06a1
#4598 MiniMax-M3.1-Flash B200 SGLang FP8 draft KV	MiniMax-M3.1	minimaxm31flash-fp4-b200-sglang-agentic-dspark (all but conc 1)	no /use yet: in progress at head 0185c537
#4524 MiniMax-M3.1-Flash B300 SGLang TP2	MiniMax-M3.1	minimaxm31flash-fp4-b300-sglang-agentic-dspark (all but conc 1)	in progress at head 3c964a22. Its /use comment names run 40345805111, which measured 6481cd6b, before the PR added FP8 draft KV.
#4509 MiniMax-M3.1-Flash B200 / GB200 / B300 SGLang conc 1	MiniMax-M3.1	the three keys above (their conc-1 configurations)	in progress at head 4231e37c. Its /use comment names run 40194550246, which measured these points at 4a2f670d under separate *-low-latency keys that the PR has since folded into the main keys.
#4595 DeepSeek-V4.1-Flash GB300 Dynamo+SGLang DEP16 disagg HiCache	DeepSeek-V4.1	new dsv41flash-fp4-gb300-dynamo-sglang-agentic-dspark-disagg-hicache; dsv41flash-fp4-gb300-dynamo-sglang-agentic-agg (new image and Dynamo router, TP2 c16 removed)	no /use yet: in progress at head e480515c
#4556 DeepSeek-V4-Pro B200 Dynamo+SGLang disagg DSpark6 HiCache	DeepSeek-V4	dsv4-fp4-b200-dynamo-sglang-agentic-disagg	no /use yet: in progress at head 99b6f1bc
#4455 Qwen3.5 FP8 H200 SGLang	Qwen3.5	qwen3.5-fp8-h200-sglang-agentic-mtp, -hicache-mtp	run 40040595183, named by the PR's /use comment (head fad84b25, unchanged), verified
#4454 Muse Spark Open NVFP4 GB300 SGLang	Muse Spark Open	musespark-fp4-gb300-sglang-agentic-dflash-multinode, -dflash-nightly1019	no /use yet: in progress at head a73b274a (unchanged)
#4333 Qwen3.5 NVFP4 GB300 Dynamo+SGLang	Qwen3.5	qwen3.5-fp4-gb300-dynamo-sglang-agentic-agg, -disagg; retires -agg-pareto, -pp-pareto	run 39355392627, named by the PR's /use comment (head 02b72dc0, 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 B200 shows InferenceX #4598, not #4536. Both change the same B200 recipe independently, so they cannot be shown together. [Bug] KeyError: 'model.layers.13.mlp.experts.w2_weight_scale' #4598 adds FP8 draft KV to main's configurations (InferenceX #4490), so [Bug] bench_speculative.py got error #4536's DRAM-offload retune and TP2 points leave the page for now. On B200, GB200 and B300, all configurations except Internvl and MMMU dataset evaluation #4509's conc 1 now come from these FP8 draft-KV PRs (in progress), no longer from their measured main runs. On GB300, InferenceX #4600 replaces the measured #4511 / #4525 curve.
• Retired keys. InferenceX #4333 deletes qwen3.5-fp4-gb300-dynamo-sglang-agentic-agg-pareto and -pp-pareto, as in [Doc] cookbook: Refresh SemiAnalysis AgentX launches to InferenceX main 67de4b34 #46948.
• 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 253c38bf. Later merges are left for the next refresh.

What this PR does

Page	Lanes	Deployments	Points	Hardware
MiniMax-M3.1	4	17	59 (was 61)	GB300, GB200, B300, B200
GLM-5.2	14	34	75	GB300, GB200, B300, B200, H200
GLM-5.3	1	1	10	B300
DeepSeek-V4	12	27 (was 26)	75 (was 70)	GB300, GB200, B300, B200, H200
DeepSeek-V4.1	10 (was 9)	24 (was 22)	69 (was 67)	GB300, GB200, B300, B200, H200, H100
Qwen3.5	16	51	220	GB300, GB200, B300, B200, H200, H100
Qwen3.8-Flash-Next	4	5	24	GB300, B300, B200, H200
Muse Spark Open	8	11	72	GB300, GB200, B300, B200, H200

• New settings kept as submitted (InferenceX #4595, DeepSeek-V4.1 GB300). Worker flags: --enable-eplb, --eplb-rebalance-num-iterations, --expert-distribution-recorder-buffer-size, --enable-fp32-lm-head and --enable-encoder-swa-bounded-replay. Dynamo workers also get DYN_BACKEND_ADMISSION_ENGINE_REQUEST_LIMIT, DYN_BACKEND_ADMISSION_ENGINE_REQUEST_WAITING_LIMIT and DYN_BACKEND_ADMISSION_REQUEST_QUEUE_LIMIT (dropped for the derived SGLang-router variant).
• MiniMax-M3.1-Flash B200 / GB200 / GB300 add --speculative-draft-kv-cache-dtype fp8_e4m3.

Validation

• Every page in a headless browser (mint dev, Mintlify as in CI). Every hardware / checkpoint / deployment / concurrency / KV tier / router combination renders exactly the decoded commands: 3,234 command sets, 0 mismatches, 0 unseen, no browser errors. Every tab was clicked, and every "Copy link" URL was reopened. The section links from the Deployment panels and the copy buttons were also checked.
• The component's own decoder reproduces every generated plan (3,234 router × KV-tier variants). Every KV flag exists in its image's SGLang commit (261 image × flag pairs). The round-trip / leak check reports 0 problems.
• Arg-parse in the submission images: all 135 commands not validated in an earlier round parse with their image's own parsers (dev-minimax-m31-flash: 102, nightly-dev-cu13-20261027-5884e4fe: 13, nightly-dev-cu13-20261028-586355a8: 20). The amd64 and arm64 builds of both nightlies come from the same SGLang commit.
• 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.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ✅ Run #40658193630
Latest PR Test (Extra): ❌ Run #40658193430
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

[mintlify]

✅ Preview approved

@nvpohanh approved a Mintlify preview for commit 9d3a4ca.

[mintlify]

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

Project	Status	Preview	Updated
sglang-doc	🟢 Ready	View Preview	Oct 28, 2026, 8:46 AM

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

[nvpohanh]

preview looks good.