[Docs] Cookbook: AgentX launches for Qwen3.5 / Qwen3.8-Flash-Next; refresh GLM-5.2, DeepSeek-V4, DeepSeek-V4.1

[nvpohanh]

[by Claude Code]

Motivation

#42726 added the interactive Agentic Long-Context Deployment explorer (docs/src/snippets/_agentx.jsx) and mirrored the SemiAnalysis InferenceX AgentX submissions for GLM-5.2, DeepSeek-V4 and DeepSeek-V4.1 into it. This PR adds the explorer to Qwen3.5-397B-A17B and Qwen3.8-Flash-Next, and refreshes GLM-5.2, DeepSeek-V4 and DeepSeek-V4.1 against the latest submissions. Each page covers every NVIDIA sglang / dynamo-sglang AgentX lane for its model, as raw launch commands linked to the run that measured them.

Source

InferenceX main (ceb50ba2), with these open InferenceX PRs overlaid:

PR	Page	Lane(s)	Pinned to
#3816 Qwen3.8 GB300 SGLang	Qwen3.8-Flash-Next	qwen3.8next-fp4-gb300-sglang-agentic-mtp	run 37714835620 (/use), verified
#3840 Qwen3.5 FP8 GB300 Dynamo+SGLang	Qwen3.5	qwen3.5-fp8-gb300-dynamo-sglang-agentic-agg-midcurve, -pp-midcurve	no /use yet: in progress, PR head 5948cad5
#2978 Qwen3.5 B300 Dynamo+SGLang	Qwen3.5	qwen3.5-fp4-b300-dynamo-sglang-agentic-pareto	run 34548026742 (/reuse-sweep-run), verified
#3817 DSV4 B300 Dynamo+SGLang	DeepSeek-V4	dsv4-fp4-b300-dynamo-sglang-agentic-agg, -disagg	run 37715955859 (/use), verified
#3606 DSV4 B300 SGLang	DeepSeek-V4	dsv4-fp4-b300-sglang-agentic-hicache-mtp	no /use yet: in progress, PR head 5b394d39
#3701 DSV4 GB200 Dynamo+SGLang	DeepSeek-V4	dsv4-fp4-gb200-dynamo-sglang-agentic-disagg	no /use yet: in progress, PR head d9185455
#3782 DSV4.1 GB300 SGLang	DeepSeek-V4.1	dsv41flash-fp4-gb300-sglang-agentic-dspark	no /use yet: in progress, PR head f00c81ab
#3744 DSV4.1 B300 SGLang	DeepSeek-V4.1	dsv41flash-fp4-b300-sglang-agentic-dspark	no /use yet: in progress, PR head ec20ed56
#3836 GLM-5.2 FP8 B300 SGLang	GLM-5.2	glm5.2-fp8-b300-sglang-agentic-mtp	no /use yet: in progress, PR head 03ffb35d

A lane from a PR without an authorized /use run is rendered at the PR head and labeled "In progress in SA InferenceX AgentX"; it switches to the measured run on the next sync after the /use comment or the merge. The overlays #42726 used (#3598, #3187, #3745, #3659) have since merged into InferenceX main; they are mirrored from there, at the same runs. #3817 replaces #3631.

What this PR does

Page	Lanes	Deployments	Points	Hardware
Qwen3.5 (new section)	19	50	221	GB300, GB200, B300, B200, H200, H100
Qwen3.8-Flash-Next (new section)	4	4	22	GB300, B300, B200, H200
GLM-5.2 (refreshed)	15	24	57	GB300, GB200, B300, B200, H200
DeepSeek-V4 (refreshed)	12	26	75	GB300, GB200, B300, B200, H200
DeepSeek-V4.1 (refreshed)	9	19	83	GB300, GB200, B300, B200, H200, H100

• Qwen3.8-Flash-Next page. An unnumbered ## Agentic Long-Context Deployment section right after the Playground, plus the opt-in agenticLink footer on its Deployment panel, as on the GLM-5.2 / DeepSeek pages.
  • The static ### 2.3 GB300 long-context serving with HiCache and MTP added in docs(cookbook): add Qwen3.8 GB300 long-context recipe #43159 is removed. Its configuration (InferenceX How to contribute an optimized R1 operator in SGlang? #3816) is now the GB300 deployment in the explorer, with the full launch command and provenance.
  • The Deployment panel's GB300 / NVFP4 (NVDA) / Low Latency cell from docs(cookbook): add Qwen3.8 GB300 long-context recipe #43159 is otherwise unchanged; only its warning's link, which pointed at the removed section, now points at the Agentic Long-Context Deployment section.
• Qwen3.5 page. This page still uses its legacy <Qwen35Deployment /> generator (no shared Deployment panel or Playground), so the explorer goes in a new ### 3.3 Agentic Long-Context Deployment at the end of ## 3. Model Deployment, with an explicit agentic-long-context-deployment anchor. The generator itself is unchanged.
  • The hand-written #### 5.2.3 Agentic Long-Context with HiCache DRAM Offload (H200 FP8, MTP) (added in Update Qwen3.5 H200 FP8 for AgentX HiCache MTP #35194) is removed. The explorer includes that lane (qwen3.5-fp8-h200-sglang-agentic-hicache-mtp) at its latest measured run. Nothing linked to the old anchor.
• GLM-5.2 / DeepSeek-V4 / DeepSeek-V4.1. Only the generated data files change; the page text is unchanged. Rendered commands change only for the overlaid lanes above, plus one derived variant:
  • [Bug] #3817 carries the DSV4 B300 Dynamo recipes from [docker] Distributed Serving with k8s Statefulset ( good example for DeepSeek-R1) #3631, with disaggregated KV transfer over NIXL (SGLANG_DISAGGREGATION_NIXL_BACKEND=LIBFABRIC) instead of Mooncake over AWS EFA. The EFA Mooncake setup step disappears from those deployments.
  • Derived HiCache variant. The DSV4 GB300 disaggregated deployments' derived HiCache variant now takes its flags from a verified B200 launch. It used to borrow them from the GB200 disaggregated lane, which is now [ROCm] Enable shfl xor sync style reduce on amd #3701's unmeasured version; only measured lanes donate KV settings.
• Component. Three small additions to _agentx.jsx; nothing else in the component, the Deployment panel or the Playground changes (apart from the one link above):
  • Deployments whose submission used --mamba-scheduler-strategy get a note that newer SGLang spells it --mamba-radix-cache-strategy (refactor: mamba radix cache server args initialize #28151), like the existing --enable-dp-attention note.
  • Deployments with SGLANG_DISAGGREGATION_NIXL_BACKEND=LIBFABRIC get a note that it reflects SemiAnalysis's AWS-EFA cluster and that InfiniBand uses NIXL's default UCX backend.
  • An install line on a derived Dynamo variant reads "This image does not bundle ai-dynamo. Install the release built for its SGLang version first: ..." instead of "Dynamo install used by SA" (see below).

How the commands are produced

Same pipeline and rules as #42726. Each lane is rendered at the exact commit of the run that produced its dashboard data or, for an open PR, the run named by its authorized /use comment (the PR head when there is none). SA-only settings are removed (benchmark harness, cluster paths and caches, telemetry, InferenceX's synthetic speculative acceptance, which is disclosed per deployment). Placeholders replace IPs, ports and NICs. Router and KV-offload variants that were not run are derived and labeled "Derived — not run". What is new here:

• Single-node lanes from job logs. The single-node Qwen lanes on main were launched by InferenceX bash launchers at their measured commits. Their commands come from the Actions job logs and a dry run of the pinned launcher, accepted only when the dry-run server command equals the logged one.
• Same shape, measured twice at one concurrency. A few Qwen lanes measured one shape at the same concurrency both with and without HiCache (for example B200 TP4: up to c64 without offload, c64–c76 with HiCache), or with different --max-running-requests caps (GB300 agg "pareto" lanes). Those become separate deployments, labeled "measured with HiCache" / "measured with no KV offload" or by the cap. Shapes without such a collision stay one deployment, as before.
• KV offload. HiCache variants borrow flags from the closest verified HiCache launch of the same model on the same hardware (an absolute --hicache-size is only reused there). Qwen3.8-Flash-Next's only HiCache launch is the GB300 one, so its B300, B200 and H200 deployments offer GPU KV only. The Mooncake variant reuses the verified DeepSeek-V4 GB300 external-linker spec on every page, and is disabled where the image's SGLang build predates --enable-unified-cache-external-linker.
• Derived Dynamo on an image without ai-dynamo. Derived Dynamo variants use the ai-dynamo bundled in the image. lmsysorg/sglang:v0.5.16-cu130 (the Qwen3.5 B200 NVFP4 and B300 single-node lanes, all submitted with the SGLang router) ships without it, so for those 10 deployments the derived Dynamo variant installs ai-dynamo 1.4.2 first. 1.4.x is the release line whose [sglang] extra pins sglang==0.5.16, and 1.4.2 is the wheel the verified Qwen3.5 GB200 Dynamo launch used.
• Not mirrored. qwen3.5-fp8-h100-sglang-agentic (H100 without MTP) has never produced dashboard data, so there is no run to pin it to.
• Images. Several nightly tags these runs used have since been pruned from Docker Hub; the page flags them and links the SGLang commit they were built from. The Qwen3.8-Flash-Next B300, B200 and H200 lanes ran lmsysorg/sglang:qwen38flashnext, a moving tag; its GB300 lane pins the arm64 build of lmsysorg/sglang:v0.5.21-cu130 by digest. [Bug] Model Stuck at Prefill and then throw "Watchdog Timeout" Error After Idle Period (Deepseek-r1:671b on two H100*8) #3836 pins its image by digest alone (lmsysorg/sglang@sha256:2d53500c…, built from SGLang 20ca564b).

The data files are generated; please review the page text and the component changes rather than the JSON.

Validation

• mint validate (strict build) and mint broken-links --check-anchors --check-redirects pass; node docs/scripts/check_cookbook_configs.mjs passes; pre-commit run --all-files passes.
• Headless-browser click-through under mint dev of every hardware / checkpoint / deployment / concurrency / KV tier / router combination on all five pages (2,328 command sets: GLM-5.2 282, DeepSeek-V4 424, DeepSeek-V4.1 498, Qwen3.5 1,048, Qwen3.8-Flash-Next 76), opening every tab. The rendered commands match the expected text exactly. For every deployment, a non-default selection's "Copy link" URL was reopened in a fresh tab and restored the same selection and tab (123 links).
• The Deployment-panel footer link appears on GLM-5.2, DeepSeek-V4, DeepSeek-V4.1 and Qwen3.8-Flash-Next, under the Playground link, and scrolls to the section without changing the Deployment selection in the URL hash. Qwen3.5's #agentic-long-context-deployment anchor sits directly above the explorer. Pages that use the Deployment panel without the prop (PaddleOCR-VL, Unlimited-OCR) still show no link.
• The component's decoder and KV-tier transform, run through Node, reproduce the Python-generated commands for all 2,328 combinations. Every KV-cache flag in a shown command exists in the SGLang commit of that deployment's image (188 image/flag pairs).
• Compared with main, GLM-5.2, DeepSeek-V4 and DeepSeek-V4.1 render byte-identical commands for every deployment outside the overlaid lanes, except the 8 derived HiCache variants described above.
• Single-node commands from bash launchers were accepted only where the dry run of the pinned launcher reproduced the server command in the run's job log (248 points).
• Every submitted and router-derived sglang.launch_server, dynamo.sglang, sglang_router.launch_router and dynamo.frontend command (1,947 distinct, across 28 images) parses with the argument parsers of its own submission image (CPU only). The 220 commands that are new in this revision were parsed now; the rest are identical to commands parsed in [Docs] Cookbook: interactive SemiAnalysis InferenceX AgentX launches for GLM-5.2, DeepSeek-V4 and DeepSeek-V4.1 #42726 or earlier in this PR.
  • Pruned nightly tags were checked with SGLang's parser from the image's exact commit, on an available nightly image.
  • Arm64-only digests were checked with the amd64 build of the same tag and SGLang commit.
  • Dynamo-specific flags were checked against the Dynamo build each run used (wheel, source commit, or the image's bundled release). The derived Dynamo variants on v0.5.16-cu130 were checked against ai-dynamo 1.4.2.

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 #37889424204
Latest PR Test (Extra): ❌ Run #37889424076
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

[nvpohanh]

Verified locally. Merge for now to unblock SA AgentX submission.