[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) 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. Syncs to InferenceX main (746150a2). Since [Doc] cookbook: Add GLM-5.3 AgentX section and refresh with eight open InferenceX PRs #45513, main merged the four Muse Spark Open retunes and the GB300 Dynamo+SGLang lane, which are now measured. It also merged image updates (mostly to v0.5.21-cu130 / v0.5.22-cu130) for 14 mirrored lanes on DeepSeek-V4, DeepSeek-V4.1, GLM-5.2, Qwen3.5 and Qwen3.8-Flash-Next.
2. Overlays five open InferenceX PRs (table below): a GLM-5.2 FP8 B200 retune with TP8 + DCP8 points, a DeepSeek-V4.1-Flash GB300 disaggregated topology and caching update, and three PRs carried over from [Doc] cookbook: Add GLM-5.3 AgentX section and refresh with eight open InferenceX PRs #45513 that now have /use runs.
3. Fixes three pull/XXX provenance links (DeepSeek-V4.1 H200 since [Doc] cookbook: Add GLM-5.3 AgentX section and refresh with eight open InferenceX PRs #45513, Muse Spark Open B200 / B300 new). Their InferenceX changelog entries were merged with a placeholder PR link; the links now name the PR that the measured commit belongs to.

Source

InferenceX main (746150a2), with these open InferenceX PRs overlaid:

PR	Page	Lanes	Pinned to
#4389 GLM-5.2 FP8 B200 SGLang retune	GLM-5.2	glm5.2-fp8-b200-sglang-agentic-mtp	no /use yet: in progress at head e390de59
#4303 DeepSeek-V4.1-Flash GB300 Dynamo+SGLang disagg	DeepSeek-V4.1	dsv41flash-fp4-gb300-dynamo-sglang-agentic-dspark-disagg	no /use yet: in progress at head a0e5be74
#4346 DeepSeek-V4 NVFP4 B300 Dynamo+SGLang, Mooncake EFA KV transfer	DeepSeek-V4	dsv4-fp4-b300-dynamo-sglang-agentic-agg, -disagg	run 39534932794, named by the PR's /use comment (head 566c6e10, same as in #45513), verified
#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, same as in #45513), verified
#3930 GLM-5.3 NVFP4 B300 SGLang MTP	GLM-5.3	glm5.3-fp4-b300-sglang-agentic-mtp	run 39551262792, named by the PR's /use comment (head c093d205), 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.

• Two merged lanes are pinned to their PRs' /use runs. InferenceX #4176 (DeepSeek-V4 B200 Dynamo+SGLang disagg) and InferenceX #4227 (GLM-5.2 GB200 Dynamo+SGLang MTP agg) merged on 2026-10-21, and their merge ingest wrote their rows. The dashboard nevertheless shows only the sibling key's run for these lanes (InferenceX #4203 agg and InferenceX #4277 disagg, ingested later the same day). Both PRs have an authorized /use run (38925477096 and 38997451681), so the lanes are pinned to those runs. On the dashboard they still appear with their sibling's run.
• InferenceX main was frozen for this PR at 746150a2. Later merges are left for the next refresh.

What this PR does

Page	Lanes	Deployments	Points	Hardware
GLM-5.2	16	32 (was 31)	74 (was 71)	GB300, GB200, B300, B200, H200
GLM-5.3	1	1	10 (was 8)	B300
DeepSeek-V4	12	25	68	GB300, GB200, B300, B200, H200
DeepSeek-V4.1	9	22 (was 21)	67 (was 66)	GB300, GB200, B300, B200, H200, H100
Qwen3.5	17	49 (was 50)	219 (was 221)	GB300, GB200, B300, B200, H200, H100
Qwen3.8-Flash-Next	4	5	24	GB300, B300, B200, H200
Muse Spark Open	6	9	61	GB300, GB200, B300, B200, H200

Only the generated data files change. The component (_agentx.jsx), the pages' text, the Deployment panels and the Playgrounds are untouched.

Overlaid PRs

• GLM-5.2 FP8 B200 SGLang (InferenceX #4389, in progress). The image moves to nightly-dev-20261020-ab580b2e (pinned by digest). The TP8 HiCache c1–c16 grid becomes eight points:
  • c1–c12: TP8 with HiCache (write-back at c1, write-through at c4–c12);
  • c16–c24: TP8 with --dcp-size 8 (decode context parallelism, KV in HBM only) and the prefill delayer (--prefill-delayer-max-delay-ms, --prefill-delayer-queue-min-ratio).
• DeepSeek-V4.1-Flash GB300 Dynamo+SGLang disagg (InferenceX #4303, in progress). Replaces the InferenceX #4180 revision:
  • prefill topology 3P1D at c1228 and 5P1D at c2884 (was 4P1D / 6P1D), plus a new 8P1D deployment at c4096;
  • prefill HiCache replaced by the Mooncake external linker on both roles (mooncake_master --client_ttl added);
  • decode round-robin routing, ZMQ event plane (DYN_EVENT_PLANE=zmq) and --disaggregation-decode-polling-interval 8;
  • image nightly-dev-cu13-20261021-c1c287a4.
• DeepSeek-V4 B300 Dynamo+SGLang (InferenceX #4346), Qwen3.5 GB300 Dynamo+SGLang (InferenceX #4333). Same commands as in [Doc] cookbook: Add GLM-5.3 AgentX section and refresh with eight open InferenceX PRs #45513, now verified by their /use runs.
• GLM-5.3 B300 (InferenceX #3930). The new /use run adds c1 and c4 to the TP8 deployment (c1–c28).

From InferenceX main

• Muse Spark Open. The GB300, GB200, B300 and B200 SGLang retunes and the GB300 Dynamo+SGLang lane (InferenceX #4342, #4343, #4341, #4350, #4344) merged. The commands are unchanged from [Doc] cookbook: Add GLM-5.3 AgentX section and refresh with eight open InferenceX PRs #45513, and the lanes are now measured by their merged runs.

• Image updates (each lane re-pinned to the run that measured it):

  Page	Lane	Image	InferenceX PR
  DeepSeek-V4	B200 Dynamo+SGLang agg	v0.5.22-cu130	#4203
  DeepSeek-V4	B200 Dynamo+SGLang disagg	v0.5.21-cu130	#4176 (/use run, see above)
  DeepSeek-V4	GB200 Dynamo+SGLang agg	v0.5.21-cu130	#4196
  DeepSeek-V4.1	B200 SGLang	nightly-dev-cu13-20261021-c1c287a4	#4365
  DeepSeek-V4.1	H100 SGLang DPA	v0.5.22-cu130	#4324
  GLM-5.2	B200 Dynamo+SGLang agg	v0.5.22-cu130	#4370
  GLM-5.2	GB200 Dynamo+SGLang disagg	v0.5.22-cu130	#4277
  GLM-5.2	GB200 Dynamo+SGLang MTP agg	v0.5.22-cu130	#4227 (/use run, see above)
  Qwen3.5	GB200 Dynamo+SGLang MTP	v0.5.22-cu130	#4278
  Qwen3.5	FP8 B200 Dynamo+SGLang disagg MTP	v0.5.22-cu130	#4235
  Qwen3.5	FP8 B200 SGLang MTP	v0.5.21-cu130	#4058
  Qwen3.5	FP8 B300 Dynamo+SGLang disagg	v0.5.22-cu130	#4224
  Qwen3.5	FP8 H200 SGLang HiCache MTP	v0.5.21-cu130	#3869
  Qwen3.8-Flash-Next	B300 SGLang MTP	v0.5.21-cu130	#4069

  Several updates also rename options that v0.5.22 removed (e.g. --cuda-graph-max-bs → --cuda-graph-max-bs-decode), as submitted.

• Qwen3.5 FP8 H200 SGLang HiCache MTP loses its TP4·EP4 deployment. InferenceX #3869 merged on 2026-10-21 by reusing a run from 2026-10-09. That run's commit predates InferenceX #4209, which added the TP4·EP4 points shown since [Doc] cookbook: Refresh SemiAnalysis AgentX launches with twelve open InferenceX PRs #45111. The dashboard now shows [Bug] bench_serving can not bench sglang with api_key auth #3869's run for this lane (TP8 only), and the page follows it.

Derived HiCache rows. Lanes newly verified this round are closer HiCache donors for some points measured without offload: DeepSeek-V4 B300, DeepSeek-V4.1 GB300 and Muse Spark Open GB300. Those points' derived "HiCache" tier now takes the donor's flags (e.g. --hicache-ratio 2 instead of 1.02 on DeepSeek-V4.1 GB300). Measured tiers are unchanged.

Not mirrored:

• musespark-fp4-h100-sglang-agentic-dflash and qwen3.5-fp8-h100-sglang-agentic: no dashboard changelog names a run for them (as in [Doc] cookbook: Add GLM-5.3 AgentX section and refresh with eight open InferenceX PRs #45513).
• qwen3.5-fp4-gb300-dynamo-sglang-agentic-agg-pareto and -pp-pareto are deleted by InferenceX #4333 (as in [Doc] cookbook: Add GLM-5.3 AgentX section and refresh with eight open InferenceX PRs #45513).

How the commands are produced

The pipeline is the same as in #45513, with these changes:

• Merged lanes without dashboard rows. A merged key whose newest dashboard changelog has no rows is pinned to its PR's authorized, successful /use run, if that run's commit has the key. This applies to InferenceX #4176 and InferenceX #4227 above. It is the same rule already used for merges that were never ingested.
• Placeholder PR links. A dashboard changelog that links pull/XXX is credited to the InferenceX PR that the measured commit belongs to.
• Rules. These are kept as submitted:
  • worker flags --dcp-size, --prefill-delayer-max-delay-ms, --prefill-delayer-queue-min-ratio and --disaggregation-decode-polling-interval;
  • the Mooncake master flag --client_ttl;
  • the Dynamo variable DYN_EVENT_PLANE.

Each lane is still rendered at the exact commit of the run that produced its data, or at the PR head while in progress.

• Removed (SA-only settings): the benchmark harness, cluster paths and caches, telemetry, and InferenceX's synthetic speculative acceptance (disclosed per deployment).
• Replaced with placeholders: IPs, ports and NICs.
• Derived variants: router and KV-offload variants that were not run are derived and labeled "Derived — not run".

The data files are generated, so there is no hand-written code to review here.

Validation

• CI gates. These all pass:

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

• Browser click-through. A headless browser clicked through every combination of hardware, checkpoint, deployment, concurrency, KV tier and router on all seven pages under mint dev, opening every tab.

  • That is 2,764 command sets: GLM-5.2 378, GLM-5.3 60, DeepSeek-V4 358, DeepSeek-V4.1 402, Qwen3.5 1,096, Qwen3.8-Flash-Next 104 and Muse Spark Open 366.
  • The rendered commands match the expected text exactly, with no browser errors.
  • The "Copy link" URL of every deployment was reopened in a fresh tab and restored the same selection and tab (146 links).
  • The Deployment-panel footer link and section placement checks pass on every page.
  • The copy buttons work on a plain-HTTP preview.

• Decoder and KV flags. The component's own decoder, run through Node, reproduces the Python-generated commands for all 2,764 combinations, including every point's measured tier. Every KV-cache flag in a shown command exists in the SGLang commit of its deployment's image.

• Round trip and leak scan. Every flag and variable of every submitted launch is kept, rewritten or removed by a rule, and no SemiAnalysis cluster artifact appears in the shown commands.

• Diff against main's data.

  Page	Command sets
  GLM-5.2	270 identical; 64 changed, 6 removed and 44 added (InferenceX SGlang blocked！！！！ #4389 retune; image updates of three Dynamo+SGLang lanes)
  GLM-5.3	48 identical; 12 added (c1, c4)
  DeepSeek-V4	284 identical; 36 changed (B200 Dynamo+SGLang image updates; GB200 agg changes only its image note); 38 derived-HiCache rows of the B300 lanes changed (new donors)
  DeepSeek-V4.1	336 identical; 12 removed and 18 added (InferenceX [Bug] Inference gets error when the batch size is large #4303); 24 changed (H100 image update); 24 derived-HiCache rows changed
  Qwen3.5	850 identical; 234 changed and 12 added (image updates); the TP4·EP4 H200 deployment leaves (see above)
  Qwen3.8-Flash-Next	84 identical; 10 changed and 10 added (B300 image update)
  Muse Spark Open	322 identical; 44 derived-HiCache rows changed (new donors)

• Cluster environment. The table matches the worker fingerprints of measured runs on every cluster with srt-slurm lanes.

• Argument parsing. 692 distinct sglang.launch_server, dynamo.sglang, sglang_router.launch_router and dynamo.frontend command lines are new since [Doc] cookbook: Add GLM-5.3 AgentX section and refresh with eight open InferenceX PRs #45513. All parse in their own submission images:

  • v0.5.22-cu130@sha256:e3a4f060… 204/204;
  • v0.5.21-cu130@sha256:b1259f3e… 259/259;
  • nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834… 96/96;
  • nightly-dev-20261020-ab580b2e@sha256:0af411bb… 43/43;
  • dev-muse-spark-open@sha256:8cfc5a72… 30/30;
  • dev-cu13-nightly-09242@sha256:c47bbe74… 24/24;
  • v0.5.20-cu130@sha256:b27fce60… 12/12;
  • nightly-dev-20260901-07c8f729 10/10;
  • nightly-dev-20261019-3c560089 10/10;
  • nightly-dev-efa-20261019-3c560089 4/4.

  Dynamo commands of lanes that pin a Dynamo source build (b4540b87) were checked against that source tree. Each new image's amd64 and arm64 builds come from the same SGLang commit. The other 1,758 of the 2,450 command lines are identical to ones parsed for earlier cookbook PRs. Environment variables are not part of this check.

No serving or benchmark runs were performed for this PR; the numbers live in the linked SemiAnalysis runs. SGLang CI was not triggered.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ✅ Run #39731179622
Latest PR Test (Extra): ❌ Run #39731179401
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

[mintlify]

✅ Preview approved

@nvpohanh approved a Mintlify preview for commit 5f542e4.

[mintlify]

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

Project	Status	Preview	Updated
sglang-doc	🟢 Ready	View Preview	Oct 22, 2026, 6:07 AM

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

[nvpohanh]

Preview looks fine. Mering to unblock AgentX submissions.