Skip to content

[NVIDIA][AgentX] MiniMax-M3 B200 vLLM digest pin, TP4 DRAM-offload tuning and capped-active c10 entries / [NVIDIA][AgentX] MiniMax-M3 B200 vLLM 按摘要固定镜像、调优 TP4 DRAM 卸载并新增限制活跃序列的 c10 配置 - #4382

Merged
adibarra merged 3 commits into
mainfrom
minimaxm3-b200-fp4-vllm-agentic-mtp-retune
Oct 23, 2026

Conversation

@hshrivastava-droid

@hshrivastava-droid hshrivastava-droid commented Oct 21, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Pins the MiniMax-M3 B200 vLLM images by digest, retunes the TP4 concurrency 15 to 38 entries (DRAM KV offload sized to the 1,073 GB node allowance), and adds a second key for arms that cap the number of active sequences.

  • Config: minimaxm3-fp4-b200-vllm-agentic-mtp, minimaxm3-fp4-b200-vllm-agentic-mtp-c10
  • Image: vllm/vllm-openai:nightly-af7f9488c2210d67e1033ecdc845b087ee7fe92b@sha256:8e2ffbb9f866de678b22e7446903b1ee5a5766c51a948013277477c462a68a77 (first key, was tag-only), vllm/vllm-openai:nightly-ac68c3087215e0a4f3cdfa218508c6aada57235d@sha256:98adb2311118263a453ec0acd5d4b72bda3db26cecc0fe8ffc94b9ab03c951f9 (new -c10 key)
  • Tested run: 39679190495 (10 benchmark and 9 eval jobs succeeded)

minimaxm3-fp4-b200-vllm-agentic-mtp (TP4, EAGLE3 K3)

  • Kept: TP4 c15 and TP4 DRAM-offload c30, c32 and c38.
  • Retuned: c30 and c38 capture CUDA graphs up to 4096 tokens with an explicit capture-size list, and set attention-config {"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}. The DRAM KV pool is sized to the 1,073 GB node allowance.
  • Removed: TP8 c1, TP4 no-offload c1, c5, c10 and c20, and TP4 DRAM c15, c20, c25, c34, c36 and c40.

minimaxm3-fp4-b200-vllm-agentic-mtp-c10 (new)

Low-concurrency arms that cap the sequences the scheduler keeps active (max-num-active-seqs) below CONC, each with its own CUDA graph capture sizes and VLLM_FLASHINFER_ALLREDUCE_BACKEND=mnnvl:

  • TP4 c10 and TP8 c10, active sequences 1, five-token draft (K5).
  • TP4 DRAM c20, active sequences 4 and 3, K3.
  • TP4 DRAM c32, active sequences 6 and 16, K3.

Notes

  • The -c10 digest sha256:98adb231... resolves to vLLM build ac68c3087215 (2026-09-30), and the run logs report that build. The image tag was corrected to nightly-ac68c3087215... to match the digest (it previously named an older commit that predates --max-num-active-seqs); the digest, and so the image that ran, is unchanged.
  • The new and changed serving settings (active-sequence caps, capture sizes, the 1,073 GB lazy KV offload, K5 drafts and mnnvl) are documented in the vLLM MiniMax-M3 recipe (guide and command builder) by vllm-project/vllm#63163, which is merged.
AI model disclosure
  • Model/version: Claude Sonnet 5.5 (claude-sonnet-5-5)
  • Role: prepared the PR
PR checklist and change type

Type of Change

  • Bug fix
  • New feature
  • Configuration change
  • Documentation update
  • Other (please describe)

Checklist

  • I have completed the AI model disclosure and kept it current
  • I have tested my changes locally
  • I have updated documentation if necessary
  • If python -m infx.matrix.changed --base origin/main lists affected config keys, I have appended one new entry covering them to the physical end of inferencex-e2e/perf-changelog.yaml; if it lists none, I have added none. I have not edited historical entries
  • If this PR appends an inferencex-e2e/perf-changelog.yaml entry, it carries exactly one primary sweep label (a maintainer applies it on fork PRs): full-sweep-fail-fast (recommended), full-sweep-enabled, or non-canary-full-sweep-enabled. Optional modifiers all-evals, evals-only, and agentx-fast require a primary label; the last two block reuse while applied.
  • Before merging via reuse, an authorized maintainer (OWNER/MEMBER/COLLABORATOR) has commented /use <run_id> (or the legacy /reuse-sweep-run) on this PR. Do this only once there is a final full sweep that is all green with evals passing, since after this comment the primary sweep label will no longer automatically kick off new sweeps. Remove and re-add the primary sweep label to force a new sweep.
中文

配置: minimaxm3-fp4-b200-vllm-agentic-mtp、minimaxm3-fp4-b200-vllm-agentic-mtp-c10;镜像: 见上方两个 vllm/vllm-openai 摘要固定镜像。重新调优 B200 vLLM MiniMax-M3 AgentX 在 15 至 38 并发区间的点位(保留 TP4 c15 与 DRAM 卸载 c30/c32/c38,移除其余原有点位),两个镜像均按摘要固定,并新增 -c10 配置:TP4/TP8 c10(限制活跃序列为 1、K5)以及 TP4 DRAM c20(活跃序列 4 与 3)和 c32(活跃序列 6 与 16)(K3)。-c10 镜像摘要对应 vLLM 构建 ac68c3087215,标签已更正为与摘要一致(摘要不变)。新增的服务设置已由 vllm-project/vllm#63163 合入上游 vLLM recipe(指南与命令构建器)。

AI 模型使用说明:Claude Sonnet 5.5(claude-sonnet-5-5),负责准备此 PR。改动类型:配置变更。检查清单见上方,除合并前的 /use 复用评论外均已勾选。

…ning and capped-active c10 entries

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Signed-off-by: hshrivastava-droid <hshrivastava@nvidia.com>
@hshrivastava-droid
hshrivastava-droid requested a review from a team October 21, 2026 21:29
@hshrivastava-droid hshrivastava-droid added the full-sweep-enabled Full sweep with canary gate; matrix jobs run to completion despite failures label Oct 21, 2026
@github-actions

Copy link
Copy Markdown
Contributor

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 claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Beyond the inline findings, I checked two more things. First, the selectors override_tp4_c15 and override_tp4_c32_dram aren't touched in this diff, but they still exist in agentic.yaml (lines 108 and 204), so they resolve. Second, the perf-changelog.yaml change only appends one block at the end of the file, covering both config keys, and leaves the existing entries alone.

Extended reasoning...

Both checks were run against the diff: the two unchanged override selectors are present in the recipe, and perf-changelog.yaml only adds one 10-line block at the tail.

Findings marked 🟡 are optional suggestions and need no follow-up push.

Comment on lines +5550 to +5553
- { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [20], srt-recipe: agentic.yaml:override_cap4_tp4_c20_dram }
- { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [20], srt-recipe: agentic.yaml:override_cap3_tp4_c20_dram }
- { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [32], srt-recipe: agentic.yaml:override_cap6_tp4_c32_dram }
- { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [32], srt-recipe: agentic.yaml:override_cap16_tp4_c32_dram }

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔴 Two pairs of new points can't be told apart by the tooling: cap4/cap3 at c20 and cap6/cap16 at c32. Their results collide, so reuse validation and per-experiment reruns fail. The c10 key reuses TP4 with the same conc and offload settings, and only the srt-recipe override differs. Exp names (generate.py:1062) and agentic_key (results/artifacts.py:123) both ignore the override. Fix: give every point a distinct identity, for example by keeping one capped variant per (tp, conc, offload), or by adding the recipe or override to the exp name and to agentic_key.

Why this was flagged

In minimaxm3-fp4-b200-vllm-agentic-mtp-c10, nvidia-master.yaml:5550-5551 are both tp 4, conc 20, dram, vllm-simple. Lines 5552-5553 are both tp 4, conc 32, dram, vllm-simple, and so is override_tp4_c32_dram at :5531 in the main key, on the same runner cluster:b200-nscale. generate.py:1062-1066 builds exp_name only from model_code, tp, conc, offload suffix and spec, so each pair gets the same exp_name. filter_exp_names (generate.py:1437-1441) then raises "Experiment name(s) matched multiple rows" when a test-config run targets either point. For single-node runs, agentic_key (results/artifacts.py:166-180) keys on hw, model prefix, framework, precision, tp, pp, dcp, pcp, ep, dp_attention, conc and offload, not the recipe. So once a sweep finishes, validate_agentic_artifacts reports "agentic point artifacts contain N duplicate row(s)". That breaks the /use <run_id> reuse path validate_reusable_sweep_artifacts.py:59. The artifact names include RECIPE_FINGERPRINT, so uploads don't collide, but the identity check doesn't use it. On the base branch every (tp, conc, offload) point was unique.

Verification: In configs/nvidia-master.yaml, lines 5550 and 5551 differ only in srt-recipe, and so do lines 5552 and 5553. infx/matrix/generate.py:1062-1066 builds the exp_name without the override, so filter_exp_names raises "Experiment name(s) matched multiple rows". agentic_key (infx/results/artifacts.py:169-183) also ignores it, so reuse validation fails. The base had no such duplicates.

Comment on lines 5524 to +5532
agentic-coding:
- dram-utilization: 0.683
search-space:
- { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1], srt-recipe: agentic.yaml }
- { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 5, 10, 15, 20], srt-recipe: agentic.yaml }
- { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [15], srt-recipe: agentic.yaml:override_tp4_c15 }
- dram-utilization: 1.0
search-space:
- { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [30], srt-recipe: agentic.yaml:override_tp4_c30_dram }
- { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [32], srt-recipe: agentic.yaml:override_tp4_c32_dram }
- { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [38], srt-recipe: agentic.yaml:override_tp4_c38_dram }

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟡 (optional) Maintainers inherit 11 overrides in the B200 MiniMax-M3 agentic.yaml that no config runs anymore. The master config now names a selector for every point. Overrides left without a selector include override_tp8_c1, override_tp4_c1/c5/c10/c20, and the dram c15/c20/c25/c34/c36/c40 blocks. Fix: delete the overrides no config references, or keep them reachable from a search-space entry. AGENTS.md says to delete recipes and assets no active config uses rather than leave them in place.

Why this was flagged

On the base branch, both search-space groups used bare srt-recipe: agentic.yaml. With no selector, expand_variants in infx/srt_slurm/variants.py:102-104 expands every override_* block, so all of them were reachable. After this diff, nvidia-master.yaml:5527-5553 points each entry at one named override. The only ones used are override_tp4_c15, override_tp4_c30_dram, override_tp4_c32_dram, override_tp4_c38_dram and the six new cap overrides. The only two configs that use srt-recipe-dir minimaxm3/vllm/b200-fp4-mtp are at nvidia-master.yaml:5522 and :5541. That leaves 11 dead blocks in agentic.yaml: override_tp8_c1 (line 56), override_tp4_c1 (69), c5 (82), c10 (95), c20 (121), c15_dram (134), c20_dram (151), c25_dram (168), c34_dram (221), c36_dram (238) and c40_dram (274). They still look like supported benchmark points, but nothing runs or checks them, which conflicts with the AGENTS.md rule to remove unused assets.

Verification: On the base, both search-space groups used bare srt-recipe: agentic.yaml. With no selector, expand_variants (inferencex-e2e/infx/srt_slurm/variants.py) expands every override_* key. After this diff, every entry at nvidia-master.yaml:5527-5553 names exactly one override. That leaves 11 overrides in agentic.yaml that nothing selects.

Comment on lines +291 to +293
# Low-concurrency arms that cap the number of sequences the scheduler keeps active
# (max-num-active-seqs) below CONC, with a three- or five-token draft. Selected by
# name from the master config.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟡 nit (optional): The new section comments break the repo's comment policy, which bans header comments above a recipe, override or section. sweep:^# (Low-concurrency arms|TP4 DRAM concurrency 20|CONC 32 with active) Lines 291-293 say what the next overrides are and how they are selected, so they are a section header, not a technical note. Line 377 is the same kind of comment, and its "cut to" wording also describes history. Fix: delete these blocks. If a value needs a reason, keep only a one-line note on that line, for example why max-num-active-seqs sits below CONC.

Why this was flagged

AGENTS.md "Code comments (mandatory)" applies to recipes and configs. It says "Never add broad, overarching comments ... above a recipe, override, or section" and "Do not describe what the file, topology, or config is". It also bans history wording. The diff adds three such comments in agentic.yaml. Lines 291-293 begin "# Low-concurrency arms that cap ... Selected by name from the master config." Lines 333-334 begin "# TP4 DRAM concurrency 20 with 4 and 3 active sequences on the c10 key's image". Line 377 begins "# CONC 32 with active sequences capped below CONC ... the CPU KV pool is cut to the 1,073 GB node allowance." Each one describes the override group below it and does not explain a specific value. The base branch has none of them.

Verification: AGENTS.md lines 60-67 apply to recipes and say: "Never add broad, overarching comments at the top of a file or above a recipe, override, or section." The diff adds such comments in agentic.yaml at lines 291-293 and 333-334, and line 377's "cut to" is history wording. Nothing breaks at runtime; the only effect is a breach of a rule the repository wrote down.

@github-actions

Copy link
Copy Markdown
Contributor

@hshrivastava-droid

Copy link
Copy Markdown
Collaborator Author

/use 39679190495

@github-actions

github-actions Bot commented Oct 22, 2026 •

Copy link
Copy Markdown
Contributor

@hshrivastava-droid staged run 39679190495.

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

@Ankur-singh

Ankur-singh commented Oct 23, 2026 •

Copy link
Copy Markdown
Collaborator

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

  • Verified that as of the moment of typing this, this is the latest version of PR_REVIEW_CHECKLIST.md
  • Verified that the general code quality meets the InferenceX standard and does not make the code quality any worse.
  • Verified that this PR has passed PR validation. Please link to GitHub Action workflow that shows this.
  • Verified that this PR passes evals. Please link to GitHub Action workflow that shows this.
  • Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
  • Verified that every draft model and draft head is served as it ships: the draft that ships with the served checkpoint, at its stored precision, through the pinned upstream image's default handling, with the shipped and effective draft precision recorded in the additional detail section. No submission-side quantization, dtype override, checkpoint substitution, or patch may lower draft precision below that default, regardless of eval results or AL. Explicitly verified that SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is not enabled in the effective recipe, including inherited settings; enabling it is prohibited going forward, and historical runs do not grant an exception. See Draft-model precision for what counts as the default and the MLPerf comparison.
  • For agentic workloads: verified that speculative-decoding configs (EAGLE / MTP / draft models) run with simulated synthetic acceptance, with the acceptance-length value taken from the committed golden AL curve in infx/golden_al_distribution/ for that model, thinking mode, and draft length. A submission may choose any supported draft length, but it may not substitute a different acceptance target.
  • Verified against the current MODELS.md that this PR does not submit a deprecated model, scenario, or model-scenario combination.
  • Verified that the model architecture isn't changed with benchmark hacks like using --hf-overrides to skipping indexer for every x layers on models that don't natively support this. As a general rule, we won't accept optimizations that reduces the number of model architecture FLOPs. Anything that makes that same computation run faster is fair game; target/verifier FLOPs at lower precisions is fine, given that the config passes private evals, but this does not permit lowering draft-model or draft-head precision below what ships. As an general north star princple, we should only use optimizations which is used in production by customers that care about accuracy
  • For affected Dynamo TRT-LLM benchmarks, verified that iteration/KV-cache metrics are collected and published under the pinned Dynamo implementation and effective engine settings. Resolve the installed Dynamo version separately from the container tag; use supported metrics controls without requiring KV events. Record the version and configuration evidence in the additional detail section. See verifier Check 17. N/A for unaffected serving paths or frameworks without these Dynamo settings.
  • If an company claims that they support vLLM/SGLang as first class LLM inference engines on their hardware, I have verified that the respective vLLM submission made using upstream https://hub.docker.com/u/vllm docker repo, upstream SGLang https://hub.docker.com/u/lmsysorg docker repo. The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet as supported by vLLM/SGLang community maintainers
  • If an company claims that they support vLLM/SGLang as first class upstream in-tree LLM inference engines on their hardware, I have have verified that the respective vLLM/SGLang submission has been made before additional frameworks (TRT-LLM, ATOM, etc.). The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet.
  • Verified that every affected vLLM/SGLang production serving configuration in this PR, including Dynamo deployments, is documented in the official vLLM recipes and/or the SGLang cookbook, for both single-node and multi-node aggregated and disaggregated serving. The documentation covers production-useful flags, environment variables, and launch commands (including prefill, decode, and router/frontend commands for disaggregation); InferenceX-only benchmark/harness settings are excluded with reasoning in the additional detail section:
    • I linked the corresponding upstream PR in the vLLM recipe repo or SGLang cookbook source repo and verified that it is MERGED before this InferenceX PR merges. An opened, draft, or closed-without-merge upstream PR does not satisfy this requirement. If the matching recipe was already published, I linked the published recipe/cookbook page in the additional detail section below.
  • Verified that this PR does not patch the inference engine or serving stack — the pinned image must run as shipped. This covers .patch files / git apply / patch, inline patches embedded in benchmark scripts (e.g. a python3/sed heredoc that rewrites installed engine sources before serving), in-place edits of site-packages, monkey-patching, overwriting container files, and installing forked/rebuilt engine wheels on top of the pinned image. The only exception is a patch covered by a filled-out waiver at docs/waiver/<PR_NUMBER>.md — named after the PR that introduces the patch and filed in that same PR, stating what is patched, why the unmodified upstream image cannot run this benchmark, the upstream PR/issue link, and the removal plan — which I have linked below in the additional detail section.
  • If this PR uses append-only: true, verified that it only adds generated points or recipe variants inside a selected existing config/scenario and existing same-image visual curve: every previously generated point remains present with the same recipe, no prior point is removed or rerun, and every benchmark-affecting change in the complete diff can affect only the corresponding newly appended points (never an existing point), regardless of which file contains it.
  • If any of the above criteria cannot reasonably be satisfied, I have provided additional reasoning below.
  • Reported measured throughput/E2EL Pareto counts and evidence per affected curve (≥5 points strongly recommended). Below 5 or unverifiable: tag a core maintainer for review; recorded admin bypass required before merge. N/A if no curves are affected. Details.

Additional detail section:

  • Assessed head: 7ddc333596f32984e0dd653f400584bf6dd25f9c (the tested commit is fb296fcb; the only change since is the one-line -c10 image tag fix in nvidia-master.yaml, with the digest unchanged) (policy checked at InferenceX main @ f6aadb8e). Configs: minimaxm3-fp4-b200-vllm-agentic-mtp (image vllm/vllm-openai:nightly-af7f9488c2210d67e1033ecdc845b087ee7fe92b@sha256:8e2ffbb9..., now digest-pinned) and the new minimaxm3-fp4-b200-vllm-agentic-mtp-c10 (image nightly-ac68c3087215e0a4f3cdfa218508c6aada57235d@sha256:98adb231...); both upstream vllm/vllm-openai, vLLM single-node aggregated on B200 (nvidia/MiniMax-M3-NVFP4, EAGLE3 draft).
  • PR validation + evals: run 39679190495 (attempt 1, success, head_sha = fb296fcb, the tested commit in this PR; the assessed head 7ddc3335 differs only by the image tag line). 10 agentic / benchmark jobs and 9 agentic eval / jobs succeeded (0 failed, 0 cancelled); the executed points match the diff exactly: old key TP4 c15, DRAM c30/c32/c38; -c10 key TP4 c10, TP8 c10 (K5, active-seqs 1), TP4 DRAM c20 (active-seqs 4 and 3) and c32 (active-seqs 6 and 16), each on its own image digest. Evals (eval_results_all, infrastructure_success: true): minimax_m3_full em_strict 0.97 to 0.98 (n=102) on seven points, gsm8k 0.9719 (TP4 c38) and 0.9697 (c32 cap6) (n=1319). Evals ran for one variant per key/TP/concurrency group, so -c10 c20 cap3 and c32 cap16 have no eval of their own. /use 39679190495 was posted by @hshrivastava-droid (COLLABORATOR) and the bot staged it.
  • Golden AL: minimaxm3_eagle3_gqa.yaml, thinking_on (the recipe sets thinking_mode enabled): 3 draft tokens = 2.78, 5 draft tokens = 3.21. The old-key points and the -c10 DRAM points use K3 (2.78) and the two -c10 c10 points use K5 (3.21); each point's applied overrides carry rejection_sample_method: synthetic with the matching synthetic_acceptance_length, and the logged mean acceptance length is 2.76 to 2.81 (K3) and 3.17 to 3.22 (K5). I read the server logs for all 10 points.
  • Chat template: AgentX replays chat-message traces against /v1/chat/completions (agentx-standalone.md).
  • Draft precision: the EAGLE3 draft is Inferact/MiniMax-M3-EAGLE3-GQA, the draft vLLM's MiniMax-M3 recipe documents for this model (a separate repo; neither nvidia/MiniMax-M3-NVFP4 nor MiniMaxAI/MiniMax-M3 embeds a draft or MTP head), served at its stored BF16 precision.
    • Shipped: the draft repo has 3 commits (latest 96692486); model.safetensors is 6,149,986,032 B, LFS sha256 0c6d6430519dd498342ccf7c94db8f7e90ba678c8ffbad3d198b35d4a0ffc33f, identical in all three commits. config.json has torch_dtype: bfloat16, no quantization_config, architecture LlamaForCausalLMEagle3, and the safetensors header shows all 17 tensors BF16 (3.07B parameters). The target nvidia/MiniMax-M3-NVFP4 is ModelOpt mixed precision (NVFP4 routed experts, MXFP8 attention, dense MLP and shared experts, BF16 embed_tokens/lm_head/router gates). The recipe sets no draft quantization, dtype, revision or hf-overrides, so it loads the latest (weight-identical) commit.
    • Effective: with an explicit draft model, vllm/config/speculative.py builds the draft ModelConfig(dtype=<target bf16>, quantization=None); target quantization is copied onto the draft only for the mtp and dspark methods, not eagle3. Eagle3LlamaForCausalLM takes its quant_config from the draft's own config (None), so all draft linears are unquantized BF16, and the draft shares the target's BF16 embed_tokens and lm_head. The draft KV cache inherits the target's --kv-cache-dtype fp8, which Check 13 allows. The flags this PR adds (indexer_kv_dtype, minimax_m3_msa_decode_backend, cudagraph sizes, VLLM_FLASHINFER_ALLREDUCE_BACKEND) apply to the target model code only; no patch or checkpoint substitution. SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is SGLang-only and absent.
    • Runtime check (run 39679190495, all 10 server logs): Resolved architecture: LlamaForCausalLMEagle3, a single-shard 5.73 GiB draft checkpoint (matches BF16; NVFP4 or FP8 would be roughly 1.5 or 3 GiB), dtype=torch.bfloat16, and no draft quantization or dtype flag. This is a source trace plus checkpoint metadata plus logs, not a per-module runtime dtype dump; the draft revision is unpinned (harmless while the weights are identical).
  • Upstream recipe (Check 3): vllm-project/vllm#63163 is MERGED (2026-10-23, ecd1e125, on main). It adds the "B200: InferenceX AgentX aggregated (vLLM)" section and the B200 NVFP4/GQA-draft builder overrides to recipes/models/MiniMaxAI/MiniMax-M3.yaml, covering both keys of this single-node aggregated vLLM PR. -vllm-agentic-mtp: K3 GQA draft with FlashInfer draft attention and local argmax reduction, --gpu-memory-utilization 0.95, VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm, per-point --max-num-seqs and capture sizes for c15/c30/c32/c38 (explicit --cudagraph-capture-sizes lists and minimax_m3_msa_decode_backend: cutlass for c30 and c38). -vllm-agentic-mtp-c10: a per-point table of --max-num-active-seqs, draft tokens (K5 for c10, K3 for c20/c32), host KV offload and --gpu-memory-utilization (including the TP8 c10 autotune-on case), VLLM_FLASHINFER_ALLREDUCE_BACKEND=mnnvl, the draft without local argmax, and the capture-size lists (max 4096, and 8192 for c20). Host KV offload: the guide gives the qualified SimpleCPUOffloadConnector config (cpu_bytes_to_use 1073000000000, lazy_offload true) with VLLM_USE_SIMPLE_KV_OFFLOAD=1 and says not to use the builder's Simple pill as is. The earlier merged GB200 coverage ([Doc] Document MiniMax-M3 GB200 Dynamo AgentX profile vllm-project/vllm#62543) is unchanged. InferenceX-only settings excluded: concurrency lists, KV_OFFLOADING/TOTAL_CPU_DRAM_GB benchmark labels, synthetic acceptance, AIPerf/trace replay. The guide still names the -c10 image by its old tag nightly-ac7509e2 (see the image bullet below).
  • Image / engine: the pinned vLLM at af7f9488 contains every flag the first key uses (max_num_active_seqs via vllm#56758, minimax_m3_msa_decode_backend, indexer_kv_dtype, --cudagraph-capture-sizes, lazy CPU offload). The -c10 key's tag and digest disagreed, and the PR now fixes it (commit 7ddc3335): the tag named vLLM ac7509e2 (2026-08-14), which predates --max-num-active-seqs, while digest sha256:98adb231... resolves to build ac68c3087215e0a4f3cdfa218508c6aada57235d (2026-09-30), and the tested run's logs report v0.30.1rc1.dev396+gac68c3087. The tag is now nightly-ac68c3087215e0a4f3cdfa218508c6aada57235d; the digest, and so the image that ran, is unchanged. The public registry no longer serves these nightly tags, so I could not cross-check the new tag against the registry; the digest is what pins the image. The merged vLLM guide still names nightly-ac7509e2 for this key. No patch, sed or pip steps in the recipe or launcher; the pinned image runs as shipped.
  • Not deprecated (MiniMax-M3 agentic coding is active in MODELS.md); no architecture hacks; append-only N/A (no append-only: true), and the PR does remove points (below). Dynamo TRT-LLM metrics item N/A (vLLM engine).
  • Pareto (run 39679190495, raw bmk_agentic_* artifacts, request_metrics.throughput.per_gpu.total_tput_tps vs latency.e2el.p90, upper-right frontier, trusted pareto_coverage.py from main): Check 14 counts one curve per dashboard series regardless of config key or image, so all 10 points are one curve, and 5 of 10 are on the frontier: old-key TP4 c15 (22.4k tput/GPU, 10.62 s), -c10 DRAM c20 cap4 (32.4k, 16.01 s), old-key DRAM c30 (46.5k, 17.60 s), -c10 DRAM c32 cap16 (50.0k, 20.29 s) and old-key DRAM c38 (52.2k, 25.46 s). That meets the 5-point recommendation with no margin: per key alone it is 4 of 4 (old key) and 2 of 6 (-c10), and the five depend on the two -c10 DRAM points sitting between old-key points. Dominated: -c10 TP4 c10, TP8 c10, c20 cap3, c32 cap6 and old-key c32. Local computation over this run only, not the dashboard.
  • Removed points: the PR drops 11 of the 15 points the old key has on the dashboard (TP8 c1; TP4 no-offload c1, c5, c10, c20; TP4 DRAM c15, c20, c25, c34, c36, c40) and keeps four (c15, DRAM c30, c32, c38). I did not verify how the dashboard treats the removed rows. The -c10 key's two K3 DRAM overrides and K5 overrides replace the whole speculative-config, so they lack the base's use_local_argmax_reduction; that is an extra difference between the keys beyond the image.
  • Warnings: the TP8 c10 point has power_valid: 0 (package_recompute_invalid), though its telemetry and power_audit artifacts exist. The PR is CONFLICTING with main in perf-changelog.yaml and nvidia-master.yaml (the latter since [Klaud Cold] Add day-0 MiniMax-M3.1-Flash B200 vLLM AgentX DSpark recipe / [Klaud Cold] 新增 MiniMax-M3.1-Flash B200 vLLM AgentX DSpark 首日配方 #4426 added the minimaxm31flash-fp4-b200-vllm-agentic-dspark key next to this PR's keys; it needs a manual merge keeping both blocks, which the reuse merge script does not resolve); a conflict-resolving merge commit creates a new head, which the reused run (/use 39679190495, posted before the tag fix) is meant to cover; the tag fix did not start a new sweep. The PR description now covers both keys, the added caps, the removed points, the corrected image tag and the merged recipe ([Recipe] MiniMax-M3: add B200 NVFP4 AgentX aggregated profile vllm-project/vllm#63163). Check 16 (NVIDIA overview coverage) was not assessed.

Signed: ankur-singh

@github-actions

github-actions Bot commented Oct 23, 2026 •

Copy link
Copy Markdown
Contributor

✅✅✅ Verdict: PASS ✅✅✅

Passed and not applicable checks

✅ Check 0 (CODEOWNER): PASS — @Ankur-singh is a named owner of inferencex-e2e/configs/nvidia-master.yaml and of the srt-slurm-recipes/**/b200-*/ path; perf-changelog.yaml is catch-all only.

✅ Check 1 (Sweep in PR): PASS — In-PR commit fb296fcb has 10/10 agentic / and 9/9 agentic eval / jobs success in run 39679190495 (this PR is agentic-only, so the single-node/fixed-seq jobs are correctly skipped); head 7ddc3335 only changes the -c10 image tag text, and the digest is unchanged.

✅ Check 2 (Evals): PASS — agg_eval_all.json: 9 results, all infrastructure_success: true; minimax_m3_full em_strict 0.97 to 0.98 (n=102) on 7 points, gsm8k 0.972 and 0.970 (n=1319); same image digests as the configs (8e2ffbb9…, 98adb231…).

✅ Check 3 (Upstream recipe): PASS — vllm-project/vllm#63163 is MERGED (2026-10-23, ecd1e125). Its B200 AgentX section covers both keys: per-point max-num-seqs / active-seq caps, K3/K5 drafts with and without local argmax, capture-size lists, the msa cutlass backend, trtllm/mnnvl allreduce, gpu-mem-util, TP8 autotune, the SimpleCPUOffload 1,073 GB config and the full serve command. The guide names the -c10 image by its old tag but with the same pinned digest.

✅ Check 4 (Reuse command): PASS — /use 39679190495 posted by hshrivastava-droid (COLLABORATOR) on its own line.

✅ Check 5 (Checklist template): PASS — All current template items are present and checked; the only unchecked item is the conditional "cannot reasonably be satisfied" item, and the additional detail section records the remaining caveats.

✅ Check 6 (Upstream images / ordering): PASS — Both entries are framework: vllm with upstream vllm/vllm-openai digest-pinned images on B200; open-source engine only, so the ordering rule is not triggered.

✅ Check 7 (Deprecated models): PASS — minimaxm3 agentic coding is active in MODELS.md (only single-turn is deprecated).

✅ Check 8 (Architecture hacks): PASS — No --hf-overrides, model-config edits or layer/expert trimming; changes are scheduling, cudagraph, attention backend, allreduce and KV-offload settings.

✅ Check 9 (Chat template): PASS — AgentX replays traces against /v1/chat/completions with default-chat-template-kwargs thinking enabled.

✅ Check 10 (Engine patches): PASS — No patch, sed, pip or file-overwrite steps in the recipe or config changes.

✅ Check 11 (Golden AL): PASS — The harness injects rejection_sample_method: synthetic, verified in job logs: K3 gets synthetic_acceptance_length 2.78 and K5 gets 3.21, matching minimaxm3_eagle3_gqa.yaml thinking_on.

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

✅ Check 13 (Draft precision): PASS — The draft is Inferact/MiniMax-M3-EAGLE3-GQA (torch_dtype: bfloat16, no quantization_config). The recipe sets no draft quantization, dtype or path override. In pinned vLLM af7f9488 speculative.py, target quantization is copied only for mtp/dspark, so the eagle3 draft loads at BF16. The draft KV cache inherits fp8 from the target (allowed). SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE does not apply (vLLM).

✅ Check 14 (Pareto coverage): PASS — One curve (minimaxm3 / agentic-coding / B200 vLLM / fp4 / run 39679190495 / p90 E2EL vs tput_per_gpu) from 10 raw bmk_agentic_* artifacts; trusted pareto_coverage.py reports 5/5 frontier points (c15, c20 cap4, c30, c32 cap16, c38).

✅ Check 15 (PR description): PASS — The body's configs, digest-pinned images, kept/removed points, -c10 arms, tag correction, tested run (10 bench + 9 eval) and merged recipe link all match the pinned code and run evidence.

✅ Check 16 (NVIDIA overview): PASS — MiniMax-M3 has a vLLM checkmark in the Framework Recipes table of the NVIDIA overview.

➖ Check 17 (Metrics publication): N/A — Direct vLLM serving with no Dynamo TRT-LLM path.

Assessed commit: 7ddc333596f32984e0dd653f400584bf6dd25f9c.

…lm-agentic-mtp-retune

# Conflicts:
#	inferencex-e2e/configs/nvidia-master.yaml
#	inferencex-e2e/perf-changelog.yaml
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

full-sweep-enabled Full sweep with canary gate; matrix jobs run to completion despite failures

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

3 participants