[ispobock]

Add a Low-Latency point at conc 1 for the B200, GB200 and B300 SGLang AgentX configs, following the GB300 recipe from #4456: TP4 with the MSA-V1 MegaKernel serving batch-size-1 DSpark steps. The point is added to each SKU's existing config, and the recipe lives in its own file, so existing points are unchanged.

中文

为 B200、GB200、B300 的 SGLang AgentX 配置新增 conc 1 低延迟点，沿用 #4456 中 GB300 的设置：TP4，由 MSA-V1 MegaKernel 处理 batch size 1 的 DSpark 步骤。该点加在各 SKU 现有的配置中，recipe 单独成文件，现有的点不受影响。

AI model disclosure

Prepared with Claude Code using claude-opus-5-5. Reviewed by the human author before submission.

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
• After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.

中文

感谢你的贡献！

• **审阅：**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER，请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，再通过 Slack 联系核心维护者。必须严格遵循模板，并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have，才能触发签核验证。
• **PR 验证：**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目，并且恰好带有一个主标签时，才会运行扫描：full-sweep-fail-fast（强烈推荐；canary 加逐矩阵 fail-fast）、full-sweep-enabled（有 canary；矩阵任务在失败后继续运行）或 non-canary-full-sweep-enabled（无 canary，也无 fail-fast）。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
• **合并后：**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复；参见重新运行失败任务的说明。

[claude]

Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.

[claude]

🟡 nit (optional): the new recipes add a comment saying where the config came from, which AGENTS.md's "No history" rule bans. sweep:# Follows the GB300 Low-Latency cell The comment above args: says what the recipe was copied from. The rule says derivation belongs in git and perf-changelog. Fix: drop "Follows the GB300 Low-Latency cell:" from each of the three new recipes. Keep only the technical part, for example # MSA-V1 MonoKernel serves batch-size-1 DSpark steps; larger batches take the standard path.

Why this was flagged

The three new recipe files b200-fp4-mtp/agentic-low-latency.yaml:24, b300-fp4-mtp/agentic-low-latency.yaml:24 and gb200-fp4-mtp/agentic-low-latency.yaml:24 each add # Follows the GB300 Low-Latency cell: the MSA-V1 MonoKernel serves batch-size-1. AGENTS.md "Code comments (mandatory)" covers recipes and says: "Do not write comments that describe how the code evolved, where it was derived from, what it replaced". The opening clause describes where the recipe was derived from, not why a value was chosen. Behaviour does not change. The cost is a convention breach that leaves maintainers a stale pointer once the GB300 recipe changes.

Verification: Each new recipe (b200-fp4-mtp, b300-fp4-mtp and gb200-fp4-mtp agentic-low-latency.yaml) adds the same two lines at 24-25, just above args:, starting "# Follows the GB300 Low-Latency cell". Line 65 of AGENTS.md says: "No history. Do not write comments that describe... where it was derived from". That clause is exactly what that rule bans. No runtime behavior is affected.

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=40535321029

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40535321029

[ispobock]

/use 40194550246

[github-actions]

@ispobock staged run 40194550246.

• MiniMax-M3.1-Flash · agentic-traces · FP4 · b200_sglang · 2026-10-25
• MiniMax-M3.1-Flash · agentic-traces · FP4 · b300_sglang · 2026-10-25
• MiniMax-M3.1-Flash · agentic-traces · FP4 · gb200_sglang · 2026-10-25

This run remains available across future /stage-results requests. Staging the same run ID again updates its staged data. Staging workflow

[Oseltamivir]

As a PR reviewer and CODEOWNER, I have reviewed this and have

Ignore missing checklist, do the other checks

[github-actions]

❌❌❌ REJECTED ❌❌❌

⚠️ Pareto coverage needs additional review: @functionstackx @cquil11 @Oseltamivir @adibarra. At least 5 points per affected throughput-versus-E2EL frontier are highly recommended. Below 5, or when coverage cannot be verified, merge only with an explicit, recorded admin bypass for the assessed commit; this advisory comment does not grant or enforce a bypass.

❗❗ WARNING ❗❗ NVIDIA official recipes overview: model/framework coverage is missing or unverified. See Check 16.

@Oseltamivir The sign-off has no checklist and no additional detail section. It doesn't link a recipe, it doesn't give draft-precision evidence, and it doesn't show from server logs that the MSA MonoKernel path ran during the evals. The sweep itself is green on an in-PR commit (run 40194550246 at 4a2f670), and /use was posted for that run.

❌ Check 2 (Evals pass): FAIL — No eval is proven to have run the changed path (2c). The gsm8k accuracy is in line with peers: 0.9666 B200, 0.9620 GB200, 0.9651 B300, all ok against sibling and cohort. Each eval ran at num_concurrent: 1 with --monokernel-backend msa (run 40194550246). But a server argument and a concurrency setting are not proof that the MSA-V1 MonoKernel dispatched. The eval server-log tarballs (srt-single-node-logs.tar.gz) could not be inspected here, and the job logs show only the launch flag. Cite server-log lines showing MonoKernel dispatch for each SKU.

❌ Check 3 (Upstream recipe): FAIL — The sign-off links no merged or published recipe. The SGLang cookbook MiniMax-M3.1 page lists its Low-Latency/MonoKernel cell for B300/GB300 and B200 only, not GB200. Link the page, show that it covers TP4, DSPARK block 7 / 8 draft tokens, monokernel-backend msa, page-size 128, disable-attn-tp-gather, flashinfer_trtllm_routed and SGLANG_FORCE_COARSE_WAR_BARRIER=1, and add GB200 coverage upstream or explain why it applies.

❌ Check 5 (Latest checklist): FAIL — The sign-off contains no items from the current PR_REVIEW_CHECKLIST.md. Every template item is missing or unchecked, and "Ignore missing checklist" is not an explanation.

⚠️ Check 13 (Draft precision): WARN — Draft precision could not be verified for the 3 *-sglang-agentic-dspark-low-latency configs (bundled DSpark drafter). No draft-quantization flags are set, and SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is not in the recipe or the logged server env. The sign-off still gives no evidence of the drafter's stored precision or how the pinned image (614b1509…) handles it by default; human review needed.

⚠️ Check 14 (Pareto coverage): WARN — Each affected curve has 1/5 points from the assessed run: MiniMax-M3.1-Flash / agentic-coding / fp4 / 2026-10-25 / p90, one each for b200_sglang, gb200_sglang and b300_sglang. Each is a single conc-1 point in run 40194550246 (agentic_* artifacts at 4a2f670). Points from other dates cannot be pooled. No admin exception has been claimed for faf5de65.

⚠️ Check 15 (PR description): WARN — The body says the "MSA-V1 MegaKernel" serves batch-size-1 DSpark steps, but the recipes and the cookbook use the MSA-V1 MonoKernel (monokernel-backend: msa, e.g. b200 recipe). Correct the English and Chinese text.

⚠️ Check 16 (NVIDIA overview): WARN — MiniMax-M3.1-Flash / SGLang on B200, GB200 and B300 is missing from the NVIDIA overview, which has no MiniMax-M3.1 row. Upstream coverage is the SGLang cookbook. Ask NVIDIA to add the pair.

Passed and not applicable checks

✅ Check 0 (CODEOWNER): PASS — The changed paths are owned by NVIDIA reviewers plus @SemiAnalysisAI/core. Oseltamivir is a listed CODEOWNER with repository admin role. Core team membership was unreadable (404), which is not a failure.

✅ Check 1 (In-PR sweep): PASS — 4a2f670 (in the PR) has 3 agentic / and 3 agentic eval / jobs, all success (run 40194550246). 16e03c7 changed comments only. Merge faf5de6 brought in only docs and vLLM B200 changes, which do not reach these SGLang configs.

✅ Check 4 (Reuse command): PASS — ispobock (COLLABORATOR) posted /use 40194550246, the Check 1 run.

✅ Check 6 (Upstream image): PASS — All 3 entries are framework: sglang on lmsysorg/sglang:dev-minimax-m31-flash@sha256:614b1509…. Only open-source engine entries were added.

✅ Check 7 (Deprecated models): PASS — minimaxm31flash (added 2026-10-23) has active Agentic coding in MODELS.md, and nothing is deprecated as of 2026-10-26.

✅ Check 8 (Architecture hacks): PASS — No --hf-overrides, model-config edits or accuracy-trading levers. The MonoKernel and the kernel backends run the same computation.

✅ Check 9 (Chat template): PASS — AgentX throughput uses golden simulated acceptance (Check 11). Evals drive /v1/chat/completions.

✅ Check 10 (Engine patches): PASS — The new recipes have no setup scripts, patches or source edits, and the launch path is the same one the sibling SGLang AgentX recipes use.

✅ Check 11 (Golden AL): PASS — The connector injects SGLANG_SIMULATE_ACC_LEN=4.04 with match-expected / real-draft-token, as the job log shows. That equals minimaxm31flash_dspark.yaml thinking_on at block size 7.

➖ Check 12 (Append-only): N/A — The new changelog entry has no append-only: true.

➖ Check 17 (Metrics publication): N/A — These are direct SGLang configs with no Dynamo; enable-metrics: true is set.

✅ Check 18 (Recipe/image/sweep consistency): PASS — Each config is TP4, 4 GPUs, conc 1 in both master and recipe, and all keys generate. The image digest is already in use and pinned. No points were removed. The changelog description is accurate, and the XXX pr-link is an accepted placeholder.

Assessed commit: faf5de65ce9ac81cf36c41a2ac216c373cc3db77.

[ispobock]

MonoKernel dispatch evidence for the evals in run 40194550246, from srt-single-node-logs.tar.gz in each eval_server_logs_* artifact (logs/<node>_agg_w0.out, TP0 lines).

B300 (dsxe-sa-b300-prd0-gpu-15, artifact ...-2701c872e94532105894_lm-eval_):

L119:  [2026-10-25 13:44:31 TP0] monokernel minimax_m3_msa_v1 on sm_103: widths [8] built in 18.6 s, 2749.6 MiB; ...
L1211: [2026-10-25 13:56:08 TP0] monokernel minimax_m3_msa_v1 on sm_103 runs target_verify steps of batch size [1] x 8 tokens (1 of 8 captured decode graphs); every other step runs the standard path

B200 (im-b200-c004, artifact ...-552d11a535b6fb6365c3_lm-eval_):

L107:  [2026-10-25 13:12:37 TP0] monokernel minimax_m3_msa_v1 on sm_100: widths [8] built in 16.9 s, 2749.6 MiB; ...
L1199: [2026-10-25 13:20:52 TP0] monokernel minimax_m3_msa_v1 on sm_100 runs target_verify steps of batch size [1] x 8 tokens (1 of 8 captured decode graphs); every other step runs the standard path

GB200 (watchtower-navy-cn01, artifact ...-d98e4470e1b1c952b4d8_lm-eval_):

L103:  [2026-10-25 13:01:30 TP0] monokernel minimax_m3_msa_v1 on sm_100: widths [8] built in 20.8 s, 2749.6 MiB; ...
L1201: [2026-10-25 13:10:46 TP0] monokernel minimax_m3_msa_v1 on sm_100 runs target_verify steps of batch size [1] x 8 tokens (1 of 8 captured decode graphs); every other step runs the standard path

During the evals every TP0 Decode batch line has #running-req: 1 and cuda graph: True (B300 1271/1271, B200 1168/1168, GB200 1189/1189), so every decode step replays the batch-size-1 graph, which is the MonoKernel one.

[Oseltamivir]

As a PR reviewer and CODEOWNER, I have reviewed this and have

Ignore missing checklist, do the other checks, look at PR comments & sgl-project/sglang#46506

[github-actions]

❌❌❌ REJECTED ❌❌❌

⚠️ Pareto coverage needs additional review: @functionstackx @cquil11 @Oseltamivir @adibarra. At least 5 points per affected throughput-versus-E2EL frontier are highly recommended. Below 5, or when coverage cannot be verified, merge only with an explicit, recorded admin bypass for the assessed commit; this advisory comment does not grant or enforce a bypass.

❗❗ WARNING ❗❗ NVIDIA official recipes overview: model/framework coverage is missing or unverified. See Check 16.

@Oseltamivir Two things block this: the sign-off still has no checklist, and the linked sgl-project/sglang#46506 (merged 2026-10-26) publishes the MonoKernel Low-Latency recipe only for B300 and GB300, so B200 and GB200 have no upstream coverage. The sweep is green on in-PR commit 4a2f670 (run 40194550246), the later commits are non-material, and /use 40194550246 was posted.

❌ Check 2 (Evals pass): FAIL — This verifier could not confirm that the changed path ran (2c). The MonoKernel dispatch lines in ispobock's comment come from the eval_server_logs_* tarballs of run 40194550246. The tarballs downloaded, but the sandbox blocked extracting them, and the eval job logs show only the --monokernel-backend msa launch flag. A human reviewer should confirm the cited agg_w0.out lines for each SKU. Accuracy itself is in line with peers: gsm8k 0.9666 on B200, 0.9621 on GB200 and 0.9651 on B300, all ok against both sibling and cohort, on the PR image.

❌ Check 3 (Upstream recipe): FAIL — B200 and GB200 Low-Latency (TP4 MonoKernel) have no merged or published recipe. B300 matches exactly: the published cookbook b300 low-latency cell has the same flags and env (--monokernel-backend msa, --page-size 128, --disable-attn-tp-gather, DSPARK block 7 with 8 draft tokens, flashinfer_trtllm_routed, ALLOW_JIT=1, SGLANG_FORCE_COARSE_WAR_BARRIER=1). B200 is different: the cookbook offers only Balanced and High-Throughput, and its config comment says "GB300 / B200 get no cell yet". GB200 is not in the Deployment config's supportedHardware. In #46506's AgentX snippet, the GB200, B200 and B300 conc-1 cells come from the standard agentic.yaml with no MonoKernel. Only GB300 has a MonoKernel cell (from #4511). Add B200 and GB200 MonoKernel coverage upstream.

❌ Check 5 (Latest checklist): FAIL — The sign-off contains none of the items in the current PR_REVIEW_CHECKLIST.md. "Ignore missing checklist" is not an explanation for leaving every item out.

⚠️ Check 13 (Draft precision): WARN — Draft precision could not be verified for the three *-sglang-agentic-dspark-low-latency configs. They use the bundled DSpark drafter. The checkpoint quantization_config (mxfp8, with NVFP4 MoE) names no draft layers, and the sign-off gives no evidence of the drafter's stored precision or of how image 614b1509… handles it by default. No draft-quantization flags are set. SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE appears neither in the recipes nor in the benchmark or eval job logs. Human review needed.

⚠️ Check 14 (Pareto coverage): WARN — Each affected curve has 1 of 5 points: MiniMax-M3.1-Flash / agentic-coding / fp4 / 2026-10-25 / p90 for b200_sglang (7423 tput/GPU), gb200_sglang (7493) and b300_sglang (7497). Each is a single conc-1 point in run 40194550246 (agentic_* artifacts at 4a2f670). No admin exception has been posted for c9d58f9; admin bypass not verified.

⚠️ Check 15 (PR description): WARN — The body says the "MSA-V1 MegaKernel" serves batch-size-1 DSpark steps, but the pinned recipes use the MSA-V1 MonoKernel (monokernel-backend: msa). Correct the English and Chinese text.

⚠️ Check 16 (NVIDIA overview): WARN — MiniMax-M3.1-Flash / SGLang on B200, GB200 and B300 is missing from the NVIDIA overview. It lists MiniMax-M2, M2.5 and M3 but has no M3.1 row. Upstream coverage is the SGLang cookbook. Ask for the pair to be added.

Passed and not applicable checks

✅ Check 0 (CODEOWNER): PASS — The NVIDIA master-config and recipe paths are owned by the NVIDIA reviewers plus @SemiAnalysisAI/core, and perf-changelog.yaml falls under the catch-all. Oseltamivir is a listed CODEOWNER (collectivex/) with repository admin role. Core team membership was unreadable (404), which is not a failure.

✅ Check 1 (In-PR sweep): PASS — 4a2f670 has 3 agentic / and 3 agentic eval / jobs, all success (run 40194550246). The later commits are non-material: 16e03c7 changed comments only, merge faf5de6 brought in docs and vLLM B200 changes, and merge c9d58f9 brought in other SGLang B300/GB300 agentic.yaml points, dependabot and uv.lock changes and the pr-link fill. None of these touch the low-latency recipes or their launch path.

✅ Check 4 (Reuse command): PASS — ispobock (COLLABORATOR) posted /use 40194550246, which is the Check 1 run.

✅ Check 6 (Upstream image): PASS — All 3 entries are framework: sglang on lmsysorg/sglang:dev-minimax-m31-flash@sha256:614b1509…, and only open-source engine entries were added.

✅ Check 7 (Deprecated models): PASS — minimaxm31flash (added 2026-10-23) has agentic coding active in MODELS.md, and nothing is deprecated as of 2026-10-26.

✅ Check 8 (Architecture hacks): PASS — There are no --hf-overrides, model-config edits or accuracy-trading levers. The MonoKernel runs the same computation.

✅ Check 9 (Chat template): PASS — AgentX replays chat traces with golden simulated acceptance (Check 11), and the evals drive chat completions.

✅ Check 10 (Engine patches): PASS — The new recipes have no setup scripts, patches or source edits. No setup, patch or sitecustomize matches exist under the minimaxm31flash SGLang recipes.

✅ Check 11 (Golden AL): PASS — All 3 agentic job logs show SGLANG_SIMULATE_ACC_LEN=4.04 with match-expected / real-draft-token. That equals minimaxm31flash_dspark.yaml thinking_on at 7 speculative tokens.

➖ Check 12 (Append-only): N/A — The new changelog entry has no append-only: true.

➖ Check 17 (Metrics publication): N/A — These are direct SGLang configs with no Dynamo, and enable-metrics: true is set.

✅ Check 18 (Recipe/image/sweep consistency): PASS — Each config is TP4 on 4 GPUs at conc 1 in both master and recipe, and all keys generate at the pinned SHA. The image digest is already in use, so nothing newly referenced needs resolving. No points were removed, and the changelog description is accurate with pr-link set to 4509.

Assessed commit: c9d58f9a746071afbff639aab07143d8c1b98bc5.

[nvpohanh]

[by Claude Code] The SGLang cookbook integration of this PR's AgentX configs (minimaxm31flash-fp4-b200-sglang-agentic-dspark, minimaxm31flash-fp4-b300-sglang-agentic-dspark, minimaxm31flash-fp4-gb200-sglang-agentic-dspark) will happen in sgl-project/sglang#46807, in the Agentic Long-Context Deployment section of the MiniMax-M3.1-Flash cookbook page.

This PR's /use comment pins run 40194550246, which measured 4a2f670d8. That commit does not have this PR's configurations under these config keys (the PR moved or added them since), so the cookbook shows the recipe at this PR's head (20a306768) as "in progress". A later cookbook PR marks it measured once a run of these keys is pinned or this PR merges.

[nvpohanh]

[by Claude Code] sgl-project/sglang#46807 has been merged. This PR's AgentX configs (minimaxm31flash-fp4-b200-sglang-agentic-dspark, minimaxm31flash-fp4-b300-sglang-agentic-dspark, minimaxm31flash-fp4-gb200-sglang-agentic-dspark) are now in the Agentic Long-Context Deployment section of the SGLang MiniMax-M3.1-Flash cookbook: https://docs.sglang.io/cookbook/autoregressive/MiniMax/MiniMax-M3.1#agentic-long-context-deployment

This PR's /use comment pins run 40194550246, which measured 4a2f670d8. That commit does not have this PR's configurations under these config keys (the PR moved or added them since), so the cookbook shows the recipe at this PR's head (20a306768) as "in progress". A later cookbook PR marks it measured once a run of these keys is pinned or this PR merges.

[nvpohanh]

[by Claude Code] The SGLang cookbook update of this PR's AgentX configs (minimaxm31flash-fp4-b200-sglang-agentic-dspark, minimaxm31flash-fp4-b300-sglang-agentic-dspark, minimaxm31flash-fp4-gb200-sglang-agentic-dspark) will happen in sgl-project/sglang#46948, in the Agentic Long-Context Deployment section of the MiniMax-M3.1-Flash cookbook page. The configs have been live in that section since sgl-project/sglang#46807; sgl-project/sglang#46948 refreshes them with the pin below.

This PR's /use comment pins run 40194550246, which measured 4a2f670d8. That commit does not have this PR's configurations under these config keys (the PR moved or added them since), so the cookbook shows the recipe at this PR's head (4231e37cf) as "in progress". A later cookbook PR marks it measured once a run of these keys is pinned or this PR merges.

[nvpohanh]

[by Claude Code] sgl-project/sglang#46948 has been merged. This PR's AgentX configs (minimaxm31flash-fp4-b200-sglang-agentic-dspark, minimaxm31flash-fp4-b300-sglang-agentic-dspark, minimaxm31flash-fp4-gb200-sglang-agentic-dspark) are now in the Agentic Long-Context Deployment section of the SGLang MiniMax-M3.1-Flash cookbook: https://docs.sglang.io/cookbook/autoregressive/MiniMax/MiniMax-M3.1#agentic-long-context-deployment

This PR's /use comment pins run 40194550246, which measured 4a2f670d8. That commit does not have this PR's configurations under these config keys (the PR moved or added them since), so the cookbook shows the recipe at this PR's head (4231e37cf) as "in progress". A later cookbook PR marks it measured once a run of these keys is pinned or this PR merges.

[Oseltamivir]

/use 40535321029

[github-actions]

@Oseltamivir staged run 40535321029.

• MiniMax-M3.1-Flash · agentic-traces · FP4 · b200_sglang · 2026-10-27
• MiniMax-M3.1-Flash · agentic-traces · FP4 · b300_sglang · 2026-10-27
• MiniMax-M3.1-Flash · agentic-traces · FP4 · gb200_sglang · 2026-10-27

This run remains available across future /stage-results requests. Staging the same run ID again updates its staged data. Staging workflow

[ispobock]

/use 40535321029

[github-actions]

@ispobock staged run 40535321029.

• MiniMax-M3.1-Flash · agentic-traces · FP4 · b200_sglang · 2026-10-27
• MiniMax-M3.1-Flash · agentic-traces · FP4 · b300_sglang · 2026-10-27
• MiniMax-M3.1-Flash · agentic-traces · FP4 · gb200_sglang · 2026-10-27

This run remains available across future /stage-results requests. Staging the same run ID again updates its staged data. Staging workflow

[Oseltamivir]

As a PR reviewer and CODEOWNER, I have reviewed this and have

Ignore missing checklist, do the other checks, look at PR comments for recipe link

[github-actions]

❌❌❌ REJECTED ❌❌❌

⚠️ Pareto coverage needs additional review. At least 5 points per affected throughput-versus-E2EL frontier are highly recommended. Below 5, or when coverage cannot be verified, merge only with an explicit, recorded admin bypass for the assessed commit; this advisory comment does not grant or enforce a bypass.

@Oseltamivir The sign-off still has no checklist, and its body only points to the PR comments instead of linking the recipe. Neither can be waived here. The sweep and evals passed on in-PR commit 4a2f670 (run 40194550246), and the low-latency recipe is unchanged at 0d59a67 (same recipe fingerprints). This verifier still could not confirm from the server logs that MonoKernel dispatched during the evals.

❌ Check 2 (Evals pass): FAIL — Could not verify that the changed path ran (2c). The MonoKernel dispatch lines that ispobock cites are inside the eval_server_logs_* tarballs of run 40194550246. The tarballs downloaded, but the sandbox blocked extracting them. The eval job logs show only the launch flags. A human reviewer should confirm the agg_w0.out lines for each SKU. Accuracy is otherwise fine: gsm8k strict 0.9666 B200, 0.9621 GB200, 0.9651 B300 on the PR image, all ok against sibling and cohort. Note that at the pinned head the generator drops this point's eval because it collides with the Balanced TP4 c1 eval identity, so the newer reuse run 40535321029 has no evals.

❌ Check 3 (Upstream recipe): FAIL — The sign-off itself links no merged or published recipe; "look at PR comments for recipe link" does not meet (a). Coverage does exist: sgl-project/sglang#46948 (merged 2026-10-28) puts B200, GB200 and B300 agentic-low-latency.yaml cells in the published cookbook. Those cells carry --monokernel-backend msa, --page-size 128, --disable-attn-tp-gather, 8 draft tokens, --max-running-requests 8, ALLOW_JIT=1 and SGLANG_FORCE_COARSE_WAR_BARRIER=1. Repost the sign-off with that link in its additional detail section.

❌ Check 5 (Latest checklist): FAIL — The sign-off contains none of the items in the current PR_REVIEW_CHECKLIST.md. "Ignore missing checklist" does not explain why the items are missing.

⚠️ Check 13 (Draft precision): WARN — Draft precision could not be verified for the B200/GB200/B300 low-latency TP4 c1 points, which use the bundled DSpark drafter (block 7). The recipes set no draft-quantization or draft-dtype flags, and SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is absent from the recipes and the launch log. The sign-off gives no evidence of the drafter's stored precision or how the pinned image (614b1509…) handles it by default, so human review is needed.

⚠️ Check 14 (Pareto coverage): WARN — Three curves are affected: MiniMax-M3.1-Flash agentic FP4 on b200_sglang, gb200_sglang and b300_sglang (P90 E2EL). Only the new c1 point on each curve has verified raw evidence, so each counts 1/5 (run 40194550246, 2026-10-25). The existing points' source artifacts were not verified, so full-curve coverage is unverifiable. No admin bypass has been posted.

Passed and not applicable checks

✅ Check 0 (CODEOWNER): PASS — Oseltamivir is listed in CODEOWNERS. The changed paths are owned by the NVIDIA owners plus @SemiAnalysisAI/core and by the * catch-all. Team membership is unreadable (404) and is not treated as a failure.

✅ Check 1 (Sweep on in-PR commit): PASS — 4a2f670 has green agentic / and agentic eval / jobs (run 40194550246). Later commits are non-material: a comment-only edit, merges, and a move of the c1 point from the *-low-latency keys into the main keys. Recipe fingerprints are identical at the pinned head, golden AL 4.04 applies in both, and benchmarks reran green on 4231e37 (run 40535321029).

✅ Check 4 (Reuse command): PASS — /use 40535321029 was posted by Oseltamivir and ispobock (both COLLABORATOR). That run is on in-PR commit 4231e37.

✅ Check 6 (Upstream image): PASS — lmsysorg/sglang:dev-minimax-m31-flash@sha256:614b1509… comes from the upstream lmsysorg org, and only open-source sglang entries are added.

✅ Check 7 (Deprecated models): PASS — minimaxm31flash agentic coding is active in MODELS.md as of 2026-10-28.

✅ Check 8 (No architecture hacks): PASS — No --hf-overrides, model-config edits, or accuracy-trading levers. MonoKernel is a kernel-path optimization.

✅ Check 9 (Chat template): PASS — These are AgentX trace-replay points whose acceptance is simulated at the golden AL (Check 11), not taken from raw-completion traffic.

✅ Check 10 (No engine patches): PASS — The new recipes have no setup scripts, patches, or source edits, and the launcher path is unchanged for these configs.

✅ Check 11 (Golden AL): PASS — The launch log applies SGLANG_SIMULATE_ACC_LEN=4.04 with match-expected/real-draft-token. This equals minimaxm31flash_dspark.yaml thinking_on for DSpark block 7.

✅ Check 12 (Append-only): PASS — Each key gains exactly one TP4 c1 point (B200 and GB200 12→13, B300 9→10) from the new agentic-low-latency.yaml file, which only these points reference. No existing point is removed or re-tuned.

✅ Check 15 (PR description): PASS — The body's claims (conc 1, TP4, MSA MonoKernel at batch size 1, added to the existing keys, separate recipe file, follows the GB300 recipe from #4456) match the pinned diff.

✅ Check 16 (NVIDIA overview): PASS — The NVIDIA overview lists MiniMax-M3.1-Flash with SGLang, which covers B200, GB200 and B300.

➖ Check 17 (Metrics publication): N/A — These are direct SGLang configs with no Dynamo worker settings. enable-metrics: true is set and AIPerf requires the sglang: metric prefix.

✅ Check 18 (Recipe consistency): PASS — Recipes resolve and have no generation errors. Master TP4/c1/kv none matches the recipe's 4 GPUs, tensor-parallel-size 4 and CONC 1. The image is not newly referenced, and no points are removed or re-tuned. The changelog entry is accurate.

Assessed commit: 0d59a6706b2e170e79a84c15de29712e9ae6502e.

[nvpohanh]

[by Claude Code] The SGLang cookbook update of this PR's AgentX configs (minimaxm31flash-fp4-b200-sglang-agentic-dspark, minimaxm31flash-fp4-b300-sglang-agentic-dspark, minimaxm31flash-fp4-gb200-sglang-agentic-dspark) will happen in sgl-project/sglang#47027, in the Agentic Long-Context Deployment section of the MiniMax-M3.1-Flash cookbook page. The configs have been live in that section since sgl-project/sglang#46807; sgl-project/sglang#47027 refreshes them with the pin below.

This PR's /use comment pins run 40194550246, which measured 4a2f670d8. That commit does not have this PR's configurations under these config keys (the PR moved or added them since), so the cookbook shows the recipe at this PR's head (4231e37cf) as "in progress". A later cookbook PR marks it measured once a run of these keys is pinned or this PR merges.

[nvpohanh]

[by Claude Code] sgl-project/sglang#47027 has been merged. This PR's AgentX configs (minimaxm31flash-fp4-b200-sglang-agentic-dspark, minimaxm31flash-fp4-b300-sglang-agentic-dspark, minimaxm31flash-fp4-gb200-sglang-agentic-dspark) are now in the Agentic Long-Context Deployment section of the SGLang MiniMax-M3.1-Flash cookbook: https://docs.sglang.io/cookbook/autoregressive/MiniMax/MiniMax-M3.1#agentic-long-context-deployment

This PR's /use comment pins run 40194550246, which measured 4a2f670d8. That commit does not have this PR's configurations under these config keys (the PR moved or added them since), so the cookbook shows the recipe at this PR's head (4231e37cf) as "in progress". A later cookbook PR marks it measured once a run of these keys is pinned or this PR merges.

[github-actions]

Ingest report — ✅ Results ingested into the production database.

Ingest workflow · Merge Ingest run · Benchmark source run

Category	new	duplicate	skipped	unmapped	failed
benchmark	3	0	0	0	0
eval	0	0	0	0	0
trace	3	0	0	0	0
power	3	0	0	0	0

_{benchmarkFiles 3, rawBenchmarkRows 3, coveredByPerJobFiles 0, purgedPoints 0, failedRunRows 0, runStatsNew 0, runStatsDuplicate 0, evalSamplesNew 0, gpuMetricSamplesNew 65700, powerPublicationPoints 3, changelogEntriesWritten 1}

Counts cover only this run. Rows: benchmark_results, eval_results, agentic rows linked to a trace, gpu_metrics series. This comment updates if the ingest is rerun.

中文

Ingest 报告 — ✅ 结果已写入 production 数据库。

上表仅统计本次运行写入的数据，各行单位依次为 benchmark_results 行、eval_results 行、已关联 trace 的 agentic 行和 gpu_metrics 序列。重新运行 ingest 时本评论会原地更新。