[Doc] cookbook: Refresh SemiAnalysis AgentX launches with DeepSeek-V4 B200 DSpark6 disagg and Muse Spark GB300 points by nvpohanh · Pull Request #46807 · sgl-project/sglang

nvpohanh · 2026-10-27T09:18:01Z

[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) 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:

Adds DeepSeek-V4 B200 Dynamo+SGLang disagg 1P1D DEP8 DSpark6 HiCache deployments (InferenceX #4556) and Muse Spark Open GB300 SGLang TP8·EP8 (2 nodes) and nightly-1019 TP4·EP4 deployments (InferenceX #4454), both in progress.
Updates MiniMax-M3.1-Flash: B200 retune and TP2 (InferenceX #4536), B300 TP2 (InferenceX #4524), conc-1 low-latency points on B200 / B300 / GB200 (InferenceX #4509), and GB300 TP2 c20–c36 (InferenceX #4525, merged).
Keeps five merged Klaud-Cold image bumps that the InferenceX dashboard does not show (see below), pinned to their /use runs.
Syncs to InferenceX main (7ca692b9) and overlays the InferenceX PRs in the table below.

Source

InferenceX main (7ca692b9), with these InferenceX PRs overlaid:

PR	Page	Lanes	Pinned to
#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 `ba157eb6`
#4536 MiniMax-M3.1-Flash B200 SGLang DRAM offload retune + TP2	MiniMax-M3.1	`minimaxm31flash-fp4-b200-sglang-agentic-dspark` (its configurations)	no `/use` yet: in progress at head `42e9b43e`
#4524 MiniMax-M3.1-Flash B300 SGLang TP2	MiniMax-M3.1	`minimaxm31flash-fp4-b300-sglang-agentic-dspark` (its TP2 configuration)	no `/use` yet: in progress at head `6481cd6b` (#46547 showed it as the separate `-tp2` key at `8eddada3`)
#4509 MiniMax-M3.1-Flash B200 / GB200 / B300 SGLang conc 1	MiniMax-M3.1	the three `minimaxm31flash-fp4-*-sglang-agentic-dspark` keys (their conc-1 configurations)	in progress at head `20a30676`. 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.
#4455 Qwen3.5 FP8 H200 SGLang	Qwen3.5	`qwen3.5-fp8-h200-sglang-agentic-mtp`, `-hicache-mtp`	no `/use` yet: in progress at head `fad84b25` (unchanged)
#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`
#4417 Muse Spark Open B300 SGLang	Muse Spark Open	`musespark-fp4-b300-sglang-agentic-dflash`	no `/use` yet: in progress at head `63c4ad27` (#46547 showed `5422503f`)
#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
#4529 (merged) DeepSeek-V4-Pro GB200 Dynamo+SGLang agg image v0.5.22-cu130	DeepSeek-V4	`dsv4-fp4-gb200-dynamo-sglang-agentic-agg`	run 40279236382, named by the PR's `/use` comment (commit `4f555851`), verified
#4515 (merged) DeepSeek-V4.1-Flash GB300 Dynamo+SGLang disagg image nightly-dev-cu13-20261025	DeepSeek-V4.1	`dsv41flash-fp4-gb300-dynamo-sglang-agentic-dspark-disagg`	run 40316025285, named by the PR's `/use` comment (commit `d30fa387`), verified
#4449 (merged) GLM-5.2 GB300 Dynamo+SGLang agg image	GLM-5.2	`glm5.2-fp4-gb300-dynamo-sglang-agentic-agg`	run 40054839636, named by the PR's `/use` comment (commit `7cc07653`, same as #46547), verified
#4324 (merged) DeepSeek-V4.1-Flash H100 SGLang DPA image	DeepSeek-V4.1	`dsv41flash-fp4-h100-sglang-agentic-dspark-dpa`	run 39339355047, named by the PR's `/use` comment (commit `32753f81`, same as #46547), verified
#4203 (merged) DeepSeek-V4-Pro B200 Dynamo+SGLang agg image	DeepSeek-V4	`dsv4-fp4-b200-dynamo-sglang-agentic-agg`	run 38960572895, named by the PR's `/use` comment (commit `e82118bb`, same as #46547), 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.

Merged Klaud-Cold image bumps the dashboard does not show. Since [Doc] cookbook: Refresh SemiAnalysis AgentX launches with MiniMax-M3.1-Flash B200 / GB200 HiCache points #46547, the InferenceX dashboard no longer serves the rows of InferenceX #4203, #4324, #4449 and #4196. It also has no rows for #4529 and #4515, which merged on 2026-10-26 with successful Merge Ingest runs. Without an overlay, those lanes would fall back to older runs (e.g. DeepSeek-V4 B200 agg to InferenceX #3257 from 2026-09-22). This PR pins each to its PR's /use run instead, so the cookbook keeps the launches it already showed (Use clang format 18 in pr-test-sgl-kernel.yml #4203 / [Bug] fix gemma-2-2b-it-FP8 accuracy #4324 / feat: Add FlashMLA submodule #4449) or shows the newer image (fix typo #4529 / Create col-major and tma-aligned x_scale for deep_gemm.gemm_fp8_fp8_bf16_nt #4515). [Bug] Flashinferv0.2.2.post1 shows Unsupported max_mma_kv: 0 error on L40 , when deploying Deepseek-V2-Lite-chat with --enable-flashinfer-mla #4196's lane is now fix typo #4529's. InferenceX #4497 (DeepSeek-V4.1 GB300 agg) changes only the perf changelog, so that lane stays on InferenceX #4434's run.
MiniMax-M3.1-Flash keys changed by several overlays are composed per configuration. On B200, InferenceX #4536 retunes agentic.yaml, so it supplies all of that recipe's configurations, and InferenceX #4509 adds conc 1. On B300, the measured configurations from main (InferenceX #4492) stay, and InferenceX #4524 and Internvl and MMMU dataset evaluation #4509 add theirs. On GB200, the measured configurations from main (InferenceX #4491) stay, and Internvl and MMMU dataset evaluation #4509 adds conc 1. Each configuration links its own run and shows its own status.
MiniMax-M3.1-Flash GB300. The dashboard's curve merges two runs of 2026-10-26: InferenceX #4511 and InferenceX #4525 (append-only: true, four TP2 HiCache points at c20–c36). Each point links the run that measured it. [Feature] generate till it reaches model's context length #4511's points render identically at Move mem_state update into debug mode #4525's commit.
Merged and ingested since [Doc] cookbook: Refresh SemiAnalysis AgentX launches with MiniMax-M3.1-Flash B200 / GB200 HiCache points #46547, now from main at the same runs: InferenceX #4490, #4491, #4434, #4346. New dashboard runs on main: DeepSeek-V4.1 B200 (InferenceX #4544), DeepSeek-V4.1 GB200 (InferenceX #4522), Muse Spark Open GB200 (InferenceX #4531), Qwen3.8-Flash-Next GB300 (InferenceX #4512).
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 with MiniMax-M3.1-Flash B200 / GB200 HiCache points #46547.
Not mirrored (unmeasured, unchanged from [Doc] cookbook: Refresh SemiAnalysis AgentX launches with MiniMax-M3.1-Flash B200 / GB200 HiCache points #46547): 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 7ca692b9. Later merges are left for the next refresh.

What this PR does

Page	Lanes	Deployments	Points	Hardware
MiniMax-M3.1	4 (was 5: B300 TP2 is now part of the B300 key)	17 (was 14)	61 (was 52)	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)	76 (was 70)	GB300, GB200, B300, B200, H200
DeepSeek-V4.1	9	22	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 (was 6)	11 (was 9)	72 (was 64)	GB300, GB200, B300, B200, H200

New serving flags kept as submitted: --disable-decode-cuda-graph and --num-reserved-decode-tokens (DeepSeek-V4 B200 disagg, InferenceX #4556), --flashinfer-autotune-skip-ops and SGLANG_PARALLEL_PROMPT_ENCODE (Muse Spark Open GB300, InferenceX #4454; the variable exists in that image's SGLang, nightly-dev-20261019-3c560089).
The GLM-5.3 page changes only in its source header (InferenceX main SHA and overlay pins).

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: 3,264 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 (3,264 router × KV-tier variants). Every KV flag exists in its image's SGLang commit (253 image × flag pairs). The round-trip / leak check reports 0 problems.
Arg-parse in the submission images: all 398 commands that changed since the last validated round parse with their image's own parsers, across 9 image digests of 8 tags: dev-minimax-m31-flash, nightly-dev-20261019-3c560089, nightly-dev-20261021-c1c287a4, nightly-dev-20261026-8252e3af, nightly-dev-cu13-20261025-65633129, nightly-dev-cu13-20261026-8252e3af, v0.5.21-cu130, v0.5.22-cu130. InferenceX #4529 pins an arm64-only digest of v0.5.22-cu130 (06441be3); its 7 commands were parsed with the amd64 build of the same tag, which has the same SGLANG_BUILD_COMMIT (baa9befa).
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 #40479730339
Latest PR Test (Extra): ❌ Run #40479730397
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

… B200 DSpark6 disagg and Muse Spark GB300 points Sync the data of all pages to InferenceX main 7ca692b9 and overlay InferenceX sgl-project#4556 (DeepSeek-V4 B200), sgl-project#4536, sgl-project#4524, sgl-project#4509 (MiniMax-M3.1-Flash), sgl-project#4455, sgl-project#4333 (Qwen3.5), sgl-project#4454, sgl-project#4417 (Muse Spark Open), and the merged sgl-project#4529, sgl-project#4515, sgl-project#4449, sgl-project#4324, sgl-project#4203, whose runs the InferenceX dashboard does not show, pinned to their /use runs. MiniMax-M3.1-Flash keys changed by several overlays are composed per configuration, each with its own run and status. The GB300 lane shows both runs the dashboard merges for it (InferenceX sgl-project#4511 and the append-only sgl-project#4525). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

mintlify · 2026-10-27T09:18:06Z

✅ Preview approved

@nvpohanh approved a Mintlify preview for commit 002216d.

mintlify · 2026-10-27T11:23:41Z

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

Project	Status	Preview	Updated
sglang-doc	🟢 Ready	View Preview	Oct 27, 2026, 11:30 AM

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

nvpohanh · 2026-10-27T11:53:26Z

preview looks good. merging

nvpohanh requested review from JustinTong0323, sogalin, wisclmy0611 and zijiexia as code owners October 27, 2026 09:18

github-actions Bot added documentation Improvements or additions to documentation deepseek labels Oct 27, 2026

mintlify Bot deployed to staging - docs October 27, 2026 11:30

View deployment

nvpohanh merged commit 359bfc1 into sgl-project:main Oct 27, 2026
114 of 118 checks passed

This was referenced Nov 4, 2026

[Doc] cookbook: Refresh SemiAnalysis AgentX launches to InferenceX main 57116fff #48392

Merged

[Doc] cookbook: Refresh SemiAnalysis AgentX launches to InferenceX main 2ad2c9f9 #48634

Merged

Navigation Menu

Repository navigation

[Doc] cookbook: Refresh SemiAnalysis AgentX launches with DeepSeek-V4 B200 DSpark6 disagg and Muse Spark GB300 points - #46807

[Doc] cookbook: Refresh SemiAnalysis AgentX launches with DeepSeek-V4 B200 DSpark6 disagg and Muse Spark GB300 points#46807
nvpohanh merged 1 commit into
sgl-project:mainfrom
nvpohanh:pohanh/cookbook-agentx-refresh-1026b

nvpohanh commented Oct 27, 2026 •

edited by github-actions Bot

Loading

Uh oh!

mintlify Bot commented Oct 27, 2026 •

edited

Loading

Uh oh!

mintlify Bot commented Oct 27, 2026 •

edited

Loading

Uh oh!

nvpohanh commented Oct 27, 2026

Uh oh!

Uh oh!

This branch was successfully deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

nvpohanh commented Oct 27, 2026 • edited by github-actions Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Source

What this PR does

Validation

CI States

Uh oh!

mintlify Bot commented Oct 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

✅ Preview approved

Uh oh!

mintlify Bot commented Oct 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

nvpohanh commented Oct 27, 2026

Uh oh!

Uh oh!

This branch was successfully deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

nvpohanh commented Oct 27, 2026 •

edited by github-actions Bot

Loading

mintlify Bot commented Oct 27, 2026 •

edited

Loading

mintlify Bot commented Oct 27, 2026 •

edited

Loading