[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) 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 (39b1aab8). Since [Doc] cookbook: Refresh SemiAnalysis AgentX launches with five open InferenceX PRs #45749, main merged InferenceX #4303 (DeepSeek-V4.1-Flash GB300 Dynamo+SGLang), InferenceX #4394 (GLM-5.2 FP4 B200 SGLang image v0.5.22-cu130), InferenceX #4389 (GLM-5.2 FP8 B200) and InferenceX #3930 (GLM-5.3 B300). All four are now measured on main.
2. Overlays seven open InferenceX PRs (table below).
3. Gives colocated workers their own GPUs and ports. Some deployments run several workers on one node. The page showed them with the same --port / bootstrap / KV-event ports, so they could not run as written. [Doc] Cookbook: update GB300 DSV4.1 disagg AgentX launches from InferenceX #4303 #45495 fixed the DeepSeek-V4.1 GB300 cell by hand; this PR makes the generator do it for every such cell (details below).

Source

InferenceX main (39b1aab8), with these open InferenceX PRs overlaid:

PR	Page	Lanes	Pinned to
#4398 GLM-5.2 NVFP4 GB200 Dynamo+SGLang	GLM-5.2	glm5.2-fp4-gb200-dynamo-sglang-agentic-agg, -disagg; retires -mtp, -mtp-agg	no /use yet: in progress at head 4d4e4c73
#4401 DeepSeek-V4-Pro FP4 GB300 Dynamo+SGLang	DeepSeek-V4	dsv4-fp4-gb300-dynamo-sglang-agentic-agg, -disagg	no /use yet: in progress at head cd647ee0
#4400 Muse Spark Open B200 SGLang	Muse Spark Open	musespark-fp4-b200-sglang-agentic-dflash	no /use yet: in progress at head 11358f48
#4404 Muse Spark Open GB300 SGLang	Muse Spark Open	musespark-fp4-gb300-sglang-agentic-dflash	no /use yet: in progress at head 316905a0
#4417 Muse Spark Open B300 SGLang	Muse Spark Open	musespark-fp4-b300-sglang-agentic-dflash	no /use yet: in progress at head ecde73b0
#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 #45749), 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 #45749), 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.

• Retired keys. InferenceX #4398 deletes glm5.2-fp4-gb200-dynamo-sglang-agentic-mtp and -mtp-agg. InferenceX #4333 deletes qwen3.5-fp4-gb300-dynamo-sglang-agentic-agg-pareto and -pp-pareto, as in [Doc] cookbook: Refresh SemiAnalysis AgentX launches with five open InferenceX PRs #45749. These keys leave the page with the overlay, as they would when the PRs merge.
• Two merged lanes stay pinned to their PRs' /use runs, as in [Doc] cookbook: Refresh SemiAnalysis AgentX launches with five open InferenceX PRs #45749. The dashboard still shows the sibling key's run for DeepSeek-V4 B200 Dynamo+SGLang disagg (InferenceX #4176, /use run 38925477096) and GLM-5.2 GB200 MTP agg (InferenceX #4227; now retired by InferenceX #4398).
• Not mirrored (unmeasured, unchanged from [Doc] cookbook: Refresh SemiAnalysis AgentX launches with five open InferenceX PRs #45749): musespark-fp4-h100-sglang-agentic-dflash and qwen3.5-fp8-h100-sglang-agentic. No dashboard changelog names them.
• InferenceX main was frozen for this PR at 39b1aab8. Later merges are left for the next refresh.

What this PR does

Page	Lanes	Deployments	Points	Hardware
GLM-5.2	14 (was 16)	34 (was 32)	75 (was 74)	GB300, GB200, B300, B200, H200
GLM-5.3	1	1	10	B300
DeepSeek-V4	12	26 (was 25)	70 (was 68)	GB300, GB200, B300, B200, H200
DeepSeek-V4.1	9	22	67	GB300, GB200, B300, B200, H200, H100
Qwen3.5	17	49	219	GB300, GB200, B300, B200, H200, H100
Qwen3.8-Flash-Next	4	5	24	GB300, B300, B200, H200
Muse Spark Open	6	9	63 (was 61)	GB300, GB200, B300, B200, H200

The generated data files change, plus three small text changes in _agentx.jsx (below). The pages' text, the Deployment panels and the Playgrounds are untouched.

Overlaid PRs

• GLM-5.2 NVFP4 GB200 Dynamo+SGLang (InferenceX #4398, in progress). Image nightly-dev-20261020-ab580b2e (pinned by digest), MTP 5-1-6.
  • Agg: TP4 and TP8 (2 nodes) points.
  • Disagg: PP8 prefill (2 nodes) with TP4 or DEP8 / DEP16 decode, in 1P6D, 1P4D, 1P2D, 1P1D, 2P1D and 3P1D shapes. These replace the DEP8-prefill MTP deployments.
• DeepSeek-V4-Pro FP4 GB300 Dynamo+SGLang (InferenceX #4401, in progress). Image nightly-dev-20261019-3c560089 (pinned by digest).
  • Agg: TP4 and TP8 (2 nodes) DSpark points.
  • Disagg: PP8 prefill (2 nodes) with DEP16 / DEP12 / DEP8 decode, in 1P1D and 2P1D shapes. These replace the DEP8-prefill 1P1D–4P1D deployments.
  • Workers add --no-speculative-enable-w4a4-mxfp4-megamoe: the draft model stays on W4A8 MegaMoE while the target runs W4A4.
• Muse Spark Open B200 / GB300 / B300 SGLang (InferenceX #4400, #4404, #4417, in progress). All three move to image nightly-dev-20261021-c1c287a4.
  • B200 adds a c64 point.
  • GB300 and B300 retune their concurrency points.
• DeepSeek-V4 B300 Dynamo+SGLang (InferenceX #4346), Qwen3.5 GB300 Dynamo+SGLang (InferenceX #4333). Same /use runs as in [Doc] cookbook: Refresh SemiAnalysis AgentX launches with five open InferenceX PRs #45749. The DeepSeek-V4 B300 commands are unchanged; the Qwen3.5 GB300 1P1D / 1P2D cells get the colocated-worker ports (below).

From InferenceX main

• DeepSeek-V4.1-Flash GB300 Dynamo+SGLang (InferenceX #4303, merged). Now measured on main (run 39695060576), at the same commit [Doc] cookbook: Refresh SemiAnalysis AgentX launches with five open InferenceX PRs #45749 and [Doc] Cookbook: update GB300 DSV4.1 disagg AgentX launches from InferenceX #4303 #45495 showed.
• GLM-5.2 FP4 B200 SGLang (InferenceX #4394). Image v0.5.22-cu130.
• GLM-5.2 FP8 B200 (InferenceX #4389), GLM-5.3 B300 (InferenceX #3930). Merged with the commands [Doc] cookbook: Refresh SemiAnalysis AgentX launches with five open InferenceX PRs #45749 showed. Now verified from main.

Colocated workers

Some deployments start several workers on one node:

• DeepSeek-V4.1 GB300 disagg: two DEP2 prefill workers per 4-GPU node.
• GLM-5.2 B200 disagg: two TP4 decode workers per 8-GPU node.
• Qwen3.5 GB300 / B200 / B300 disagg: prefill and decode on one node, or two decode workers per node.

SemiAnalysis's launches gave each of them its own GPUs and ports. The page now does too:

• Several colocated workers in one role. The worker tab starts with a LOCAL_WORKER_RANK preamble. It sets CUDA_VISIBLE_DEVICES, PORT (+32 per rank, srt-slurm's spacing: SGLang derives further ports from --port), BOOTSTRAP_PORT (+1) and the KV-event endpoint. The note under the tab lists each worker's rank.
• KV-event endpoint stride. The endpoint is offset by the attention-DP size per rank, because a publisher binds endpoint + DP rank. srt-slurm's +1 spacing would overlap for DEP2; [Doc] Cookbook: update GB300 DSV4.1 disagg AgentX launches from InferenceX #4303 #45495 used the same fix.
• A single colocated worker (e.g. the decode worker of a one-node 1P1D) shows its offset ports directly, e.g. --port 30032.
• SGLang router URLs in the derived variant use the same per-worker ports.
• Placement note. These cells carry a note that the GPU count covers the GPUs the workers use, so whole nodes should be reserved.

Relation to #45495. That PR hand-edited the DeepSeek-V4.1 GB300 disagg cell. This sync replaces the edit with generated data:

• Kept: the per-worker GPU pinning, distinct ports, the 2-port KV-event stride and a Placement note.
• Changed: the cell again offers the derived no-offload / HiCache tiers next to the measured Mooncake tier, labeled "derived", like every other cell.
• Removed by existing rules: --enable-metrics, DYN_SYSTEM_PORT, PYTHONUNBUFFERED and MC_STORE_CLIENT_METRIC_INTERVAL. The Dynamo system server is off by default, so dropping DYN_SYSTEM_PORT cannot make colocated workers collide.

_agentx.jsx changes only text: the worker note for blocks with local_ranks, and the LOCAL_WORKER_RANK placeholder description (no longer prefill-only).

Flag rules added: --no-speculative-enable-w4a4-mxfp4-megamoe (keep, kernel selection) and UCX_CUDA_IPC_ENABLE_MNNVL (keep, UCX CUDA IPC over multi-node NVLink for NIXL, InferenceX #4398).

Validation

• Every page in a headless browser (mint dev, Mintlify 4.2.559 as in CI). Every hardware / checkpoint / deployment / concurrency / KV tier / router combination renders exactly the decoded commands: 2,814 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 (2,814 router × KV-tier variants). Every KV flag exists in its image's SGLang commit (218 image × flag pairs). The round-trip / leak check reports 0 problems.
• Arg-parse in the submission images: all 358 commands that changed since the last validated round parse with the image's own SGLang / Dynamo parsers. nightly-dev-20260910-00840301 (pruned) was checked with its SGLang commit's source on nightly-dev-cu13-20260901-07c8f729. The colocated workers' preamble was also run in bash.
• 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 #39916152723
Latest PR Test (Extra): ❌ Run #39916152688
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

[mintlify]

✅ Preview approved

@nvpohanh approved a Mintlify preview for commit 0f4b939.

[mintlify]

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

Project	Status	Preview	Updated
sglang-doc	🟢 Ready	View Preview	Oct 23, 2026, 7:33 AM

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

[nvpohanh]

Preview looks good. Merging to unblock AgentX submission