Skip to content

Add MiniMax M3.1 Flash GB300 Dynamo TRT AgentX configurations / 新增 MiniMax M3.1 Flash GB300 Dynamo TRT AgentX 配置 - #4751

Open
RohitNagraj wants to merge 8 commits into
mainfrom
minimaxm31flash-gb300-trt-agentx
Open

RohitNagraj wants to merge 8 commits into
mainfrom
minimaxm31flash-gb300-trt-agentx

Conversation

@RohitNagraj

@RohitNagraj RohitNagraj commented Nov 2, 2026 •

Copy link
Copy Markdown
Collaborator

Add aggregated and disaggregated GB300 AgentX configurations for MiniMaxAI/MiniMax-M3.1-Flash using Dynamo, TensorRT-LLM, DSpark, native DRAM KV offload, and NIXL transfer. Configure iteration and KV-cache statistics for each TensorRT-LLM worker, with iteration logging disabled.

Validation from the pinned sweep:

  • Sweep and accuracy: all 10 benchmark jobs and 9 eval-only jobs passed. GSM8K exact match is 96.0–97.0% on every evaluated point, in line with sibling points and the vLLM GB300 configuration.
  • Image and Dynamo: the TensorRT-LLM image resolves and ran with Dynamo 1.7.0.dev20261028.
  • Power: GPU and host CPU power collection is enabled and produced valid data for all 10 points.
  • Synthetic acceptance: benchmark jobs ran with TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS=3.04 (golden AL 4.04 minus 1) on every speculating role. Eval-only jobs ran without it.
  • Draft precision: the DSpark draft is the one shipped in MiniMaxAI/MiniMax-M3.1-Flash (dspark/), loaded with dtype='auto'. There is no submission-side quantization, dtype override or checkpoint substitution. The shipped attention and MLP projections are MXFP8 (1x32 blocks), the other draft weights are BF16, and the draft KV cache is BF16. TP4 workers log the MXFP8 drafter GEMV on 20 projections. On TP2 workers, TRTLLM_M31_SM103_TP2_TARGET=1 turns off the MXFP8 drafter GEMV; the effective drafter projection precision on that path is not yet verified.
  • Upstream recipe: not required. The recipe requirement covers vLLM and SGLang, and these configurations use TensorRT-LLM.
AI model disclosure
  • Model/version: GPT-6; the exact version identifier could not be verified from the runtime.
  • Role: Prepared the configuration and recipe changes and drafted the PR.
  • Devin (Cognition AI) updated the validation summary in this description from the pinned sweep's artifacts.
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

新增 MiniMaxAI/MiniMax-M3.1-Flash 在 GB300 上的聚合式和分离式 AgentX 配置,使用 Dynamo、TensorRT-LLM、DSpark、原生 DRAM KV 卸载及 NIXL 传输。为每个 TensorRT-LLM worker 配置迭代统计和 KV 缓存统计,并关闭逐迭代日志。

基于固定的 sweep的验证结果:

  • Sweep 与准确率: 10 个 benchmark 任务和 9 个 eval-only 任务全部通过。所有评测点的 GSM8K exact match 为 96.0–97.0%,与同组其他点及 vLLM GB300 配置一致。
  • 镜像与 Dynamo: TensorRT-LLM 镜像可解析,并与 Dynamo 1.7.0.dev20261028 一起运行。
  • 功耗: 已启用 GPU 和主机 CPU 功耗采集,10 个点均产生有效数据。
  • 合成接受率: benchmark 任务在所有推测解码角色上使用 TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS=3.04(golden AL 4.04 减 1)。eval-only 任务未设置该变量。
  • 草稿模型精度: DSpark 草稿模型即 MiniMaxAI/MiniMax-M3.1-Flash 自带的 dspark/,以 dtype='auto' 加载,没有提交侧量化、dtype 覆盖或检查点替换。出厂的 attention 与 MLP 投影为 MXFP8(1x32 块),其余草稿权重为 BF16,草稿 KV 缓存为 BF16。TP4 worker 日志显示 MXFP8 drafter GEMV 作用于 20 个投影。在 TP2 worker 上,TRTLLM_M31_SM103_TP2_TARGET=1 关闭了 MXFP8 drafter GEMV,该路径下草稿投影的实际精度尚未验证。
  • 上游 recipe: 不需要。recipe 要求仅覆盖 vLLM 和 SGLang,这些配置使用 TensorRT-LLM。

AI 使用说明:使用 GPT-6 准备配置、recipe 和 PR 草稿;运行环境未提供可确认的完整版本标识。Devin(Cognition AI)根据固定 sweep 的产物更新了本描述中的验证结果。

新增 MiniMax M3.1 Flash GB300 TRT AgentX 配置。
@github-actions

github-actions Bot commented Nov 2, 2026

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。

@RohitNagraj RohitNagraj added the non-canary-full-sweep-enabled Run the full sweep without the canary gate (full search space, no trim) label Nov 2, 2026
为 GB300 TensorRT-LLM AgentX 配置启用功耗采集,并补充缓存数据类型说明。
为 GB300 TRT AgentX 更新日志填写拉取请求链接。
@github-actions

github-actions Bot commented Nov 2, 2026 •

Copy link
Copy Markdown
Contributor

gpus_per_node: 4
engine:
type: trtllm
served_model_name: MiniMaxAI/MiniMax-M3.1-Flash

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

add

    numa_cpu_bind: true
    numa_memory_bind: preferred

env:
ETCD_LEASE_TTL: '120'
DYN_ROUTER_ACTIVE_REQUEST_EXPIRY_SECS: '3600'
DYN_TOKENIZER_CACHE_BYTES: '8589934592'

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

add

DYN_TOKENIZER: fastokens

TRTLLM_M31_NO_MIX_DECODE_INTERVAL: '1'
TRTLLM_KV_TRANSFER_NUM_THREADS: '4'
DYN_ROUTER_MODE: kv
DYN_ROUTER_SESSION_AFFINITY_TTL_SECS: '14400'

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

change

        DYN_ROUTER_MODE: kv
        DYN_ROUTER_SESSION_AFFINITY_TTL_SECS: '14400'

on decode workers to

DYN_ROUTER_MODE: least-loaded

please leave prefill worker's session affinity settings untouched.

TRTLLM_KV_ONBOARD_CTAS: '12'
TRTLLM_KV_TRANSFER_NUM_THREADS: '4'
DYN_ROUTER_MODE: kv
DYN_ROUTER_SESSION_AFFINITY_TTL_SECS: '14400'

@richardhuo-nv richardhuo-nv Nov 3, 2026 •

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

add

DYN_ENGINE_CONV_AFFINITY: '1'

on prefill workers.

RohitNagraj and others added 5 commits November 3, 2026 22:37
合并 main 分支以更新 MiniMax AgentX 配置。
调整 MiniMax AgentX 路由、分词器、NUMA 绑定及遥测配置。
合并 main 分支并解决 MiniMax GB300 配置冲突。
Add 1P4D TP4 at concurrency 40, 36 and 32, and configure 1P3D TP4 at 32, 28 and 24.

新增并行度为 40、36 和 32 的 1P4D TP4 配置,并将 1P3D TP4 的并行度设为 32、28 和 24。
更新 GB300 AgentX 配置矩阵。
@RohitNagraj

Copy link
Copy Markdown
Collaborator Author

/use 42086507846

@github-actions

github-actions Bot commented Nov 10, 2026 •

Copy link
Copy Markdown
Contributor

@RohitNagraj staged run 42086507846.

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

@kedarpotdar-nv

kedarpotdar-nv commented Nov 10, 2026 •

Copy link
Copy Markdown
Collaborator

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

Checklist items
  • 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
  • Verified that new models disable integer all-reduce, with SGLang ROCm explicitly setting ROCM_QUICK_REDUCE_QUANTIZATION=NONE. Existing models are grandfathered per MODELS.md.
  • 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
  • For official InferenceX (non-preview), verified that established models use upstream engine images built from master (or the equivalent default development branch), with new models transitioning within a reasonable time after release. See MODELS.md.
  • 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:

  • Checklist template version: 8591a7e4.
  • Validation: run 42086507846 (attempt 6) ran on head 2f8becf2. All 10 benchmark jobs and all 9 eval jobs passed. /use 42086507846 was posted by RohitNagraj (COLLABORATOR).
  • Evals: gsm8k em_strict was 0.960–0.970 at every evaluated point, with n=1319 and SE ≈0.005. The 1P4D TP4 c36 point shares the recipe of the evaluated c40 point.
  • Golden AL (agentic): the curve is minimax-m3.1-flash__dspark__native__probabilistic-block.yaml with thinking on. The golden AL at K7 is 4.04. The harness injects TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS=3.04, which is the golden AL minus 1. The AL gate on the 1P3D eval measured 3.97 against 4.04 (ratio 0.98) and passed.
  • Draft precision:
    • Shipped: the DSpark draft at MiniMaxAI/MiniMax-M3.1-Flash/dspark has MXFP8 attention and MLP projections (quant_method: mxfp8, block 32). Its other weights (fc, heads and norms) are BF16. The draft KV cache is a separate BF16 pool.
    • Effective on TP4 workers: MXFP8. The worker logs show it loading /model/dspark/model-dspark-00001.safetensors with "MXDRAFT: MXFP8 drafter", and they log MXFP8 drafter GEMV on 20 projections.
    • Effective on TP2 workers (agg c22/c32/c44/c48, 1P2D TP2 c32, 1P3D TP2 c36): BF16, which is higher than the shipped precision. TRTLLM_M31_SM103_TP2_TARGET=1 turns off the MXFP8 drafter GEMV (MXFP8 drafter GEMV on 0 projections). We checked this against the installed tensorrt_llm source in the pinned image:
      • At load, prepare_minimax_dspark_weights converts each shipped MXFP8 projection weight to BF16 exactly (dequantize_mxfp8_weight).
      • With MXDRAFT off, route_mx_linears never runs, so q/k/v/o and gate/up/down use the stock BF16 F.linear.
      • The image's drafter files match NVIDIA/TensorRT-LLM#21038.
    • TRTLLM_M31_AR_NVFP4, TRTLLM_M31_MOE_AR, DECODE_FUSION, LOWM_GEMM, FAST_SAMPLER and use_low_precision_moe_combine don't lower draft projection precision. They act only on the target's MoE path, the MXFP8 GEMM kernel choice or sampling, and the draft is dense.
    • There is no draft dtype override or substitution, and SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE does not apply.
  • All-reduce: the TP2 agg and decode workers set TRTLLM_M31_AR_NVFP4=1. That is an NVFP4 all-reduce, which is not integer-quantized and so is allowed under the all-reduce policy. The TP4 variants leave it unset. The TP2 eval points cover it.
  • KV cache: all roles use NVFP4 KV. The indexer KV is NVFP4 on TP2 and FP8 on TP4. Both are stated in the changelog.
  • Image:
    • The image is nvcr.io/nvidia/tensorrt-llm/release:feat-m31-flash-perf1-1.4.0a0.dev202611010306. The NGC manifest exists (HTTP 200).
    • It is a model-enablement image for a new model (added 2026-10-23), which the main-branch image policy allows temporarily.
    • Engine-first is satisfied: GB300 vLLM, SGLang, Dynamo-vLLM and Dynamo-SGLang MiniMax-M3.1-Flash configs are already on main.
  • Upstream recipe: N/A. This is TRT-LLM, not a vLLM or SGLang submission.
  • Patches: none. The only additions are the Dynamo frontend wheel 1.7.0.dev20261028 and the TRTLLM_M31_* runtime switches that ship in the pinned image.
  • Dynamo metrics (Check 17):
    • The installed Dynamo is ai-dynamo==1.7.0.dev20261028, resolved from the worker env and pip logs.
    • publish_metrics: true results in --publish-metrics on every worker. enable_iter_perf_stats: true is set with iter_perf_stats_interval: 20, and kv_cache_config.iteration_stats_interval is 20.
    • AIPERF_REQUIRED_SERVER_METRIC_PREFIX=trtllm_kv_cache_utilization is set, and the metric was collected.
  • Stream interval: every role sets stream_interval: 20, which meets the new AgentX cap of 20 or less.
  • Power: power_valid and cpu_power_valid are 1 on all 10 points.
  • Errors: each agg point has only 1–2 errored records, all AIPerf "no responses with actual content" errors. The larger records_error_dropped count overlaps the warmup drops. Disagg points have 0 errors.
  • Reviewer feedback: richardhuo-nv's requests are all applied: NUMA binding, fastokens, least-loaded decode routing, and prefill conversation affinity.
  • Removed points / append-only: N/A, because these are new configs.
  • Requested from the author before merge:
    • Add the NVFP4 all-reduce (TRTLLM_M31_AR_NVFP4: '1') and moe_config.use_low_precision_moe_combine: true to the perf-changelog.yaml entry.
    • Merge main to resolve the conflict.
    • The PR body has been updated: it now records power collection, accuracy and draft precision.
  • Ownership: kedarpotdar-nv covers nvidia-master.yaml and the gb300-* recipes. perf-changelog.yaml is @SemiAnalysisAI/core only.

Pareto (tput/GPU vs P90 E2EL): 7 of 10 points are on the frontier. On tput/GPU vs mean interactivity, all 10 are.

Point tput/GPU P90 E2EL (s) Mean intvty Frontier (E2EL)
Agg TP2 c22 79,979 11.51 300.5 ✅
Agg TP2 c32 113,250 14.98 226.1 ✅
Agg TP2 c44 132,803 19.70 172.6 ✅
Agg TP2 c48 140,573 22.34 150.8 ✅
1P2D TP2 c32 41,357 9.82 446.1 ✅
1P3D TP2 c36 35,234 8.46 496.4 ✅
1P3D TP4 c24 11,993 7.99 1103.2 ✅
1P2D TP4 c36 23,539 9.61 818.6 ❌ dominated by 1P2D TP2 c32
1P4D TP4 c36 13,322 15.81 1068.5 ❌ dominated by 1P2D TP2 c32
1P4D TP4 c40 13,690 19.46 1065.5 ❌ dominated by 1P2D TP2 c32

Edited by Ankur-singh with Devin, after informing kedarpotdar-nv. The edit adds the TP2 draft-precision evidence and updates the changelog request and the PR body status.

Signed: kedarpotdar-nv

@github-actions

github-actions Bot commented Nov 10, 2026 •

Copy link
Copy Markdown
Contributor

❌❌❌ REJECTED ❌❌❌

@kedarpotdar-nv Blocking: the new perf-changelog.yaml entry still leaves out the NVFP4 all-reduce quantization that the TP2 workers use. The edited sign-off lists this under "Requested from the author before merge", but it has not been added at 2f8becf2. Changelog entries are permanent, so it must be fixed before merge.

❌ Check 18 (Recipe, image, and sweep consistency): FAIL — (f2) The changelog entry leaves out a quantization change. The TP2 agg workers (c22, c32, c44) and the TP2 decode workers in 1P2D c32 and 1P3D c36 set TRTLLM_M31_AR_NVFP4: '1' (agg-variants.yaml, disagg-variants.yaml), but the PR 4751 entry in perf-changelog.yaml mentions only the NVFP4 KV and the NVFP4/FP8 indexer KV. Add the TP2 NVFP4 all-reduce to the entry, and preferably moe_config.use_low_precision_moe_combine: true as well. (a)-(e), (g) and (i) pass: both configs generate, master and recipe topology agree, the image resolves, the Dynamo wheel is pinned to 1.7.0.dev20261028, no points are removed, and every role sets stream_interval: 20.

⚠️ Check 13 (Draft precision): WARN — Draft precision could not be verified for the TP2 variants (agg TP2 c22/c32/c44/c48, 1P2D TP2 c32, 1P3D TP2 c36), which set TRTLLM_M31_SM103_TP2_TARGET=1. The edited sign-off says this path dequantizes the shipped MXFP8 projections to BF16 and logs MXFP8 drafter GEMV on 0 projections, which would be at or above the shipped precision. NVIDIA/TensorRT-LLM#21038 (mxdraft.py, at 8b1d98e9) does confirm that the stock drafter dequantizes the MXFP8 weights to BF16 and that the MXDRAFT route is the only MXFP8-GEMV path. That PR is still open, though. None of its changed file paths contains an sm_103__tp2 target, and the PR does not appear to define the recipe's TRTLLM_M31_SM103_TP2_TARGET switch. The TP2 worker logs in run 42086507846 sit inside a multinode_server_logs.tar.gz that could not be extracted here, and the job log shows only the env. The PR body also still says the TP2 drafter precision is "not yet verified". A human reviewer needs to confirm the TP2 log line, or the pinned image's TP2 path, before merge. SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is not set anywhere in this TRT-LLM recipe.

Passed and not applicable checks

✅ Check 0 (CODEOWNER): PASS — kedarpotdar-nv is a listed owner of inferencex-e2e/configs/nvidia-master.yaml and of srt-slurm-recipes/**/gb300-*/. perf-changelog.yaml falls only under the catch-all * @SemiAnalysisAI/core, which any recognized CODEOWNER satisfies.

✅ Check 1 (Sweep and evals on an in-PR commit): PASS — Run 42086507846 ran on the pinned head 2f8becf2, which is still the PR tip. All 10 multi-node agentic / and all 9 multi-node agentic eval / check-runs succeeded, and no commits come after it.

✅ Check 2 (Eval accuracy): PASS — All 9 gsm8k rows ran on feat-m31-flash-perf1-1.4.0a0.dev202611010306, with em_strict 0.960–0.970 (n=1319). Every row is within 2σ of its best sibling and of the ingested vLLM GB300 DSpark cohort (run 41552552280, 0.963/0.966). Both the TP2 lanes (NVFP4 all-reduce, low-precision MoE combine) and the TP4 lanes are covered.

➖ Check 3 (Upstream recipe): N/A — Both affected configs are dynamo-trt (TensorRT-LLM), so no vLLM or SGLang production serving configuration is affected.

✅ Check 4 (Reuse command): PASS — /use 42086507846 was posted on its own line by RohitNagraj (COLLABORATOR) on 2026-11-10, naming the Check 1 run.

✅ Check 5 (Checklist template): PASS — All 20 items of the current PR_REVIEW_CHECKLIST.md, including the recipe sub-item, are present and checked.

✅ Check 6 (Upstream images and engine-first): PASS — (a) does not apply because the framework is TRT-LLM. (b) GB300 MiniMax-M3.1-Flash vllm, sglang, dynamo-vllm and dynamo-sglang entries already exist in nvidia-master.yaml. (c) The image is a model-enablement build for a newly released model, which the policy temporarily allows.

✅ Check 7 (Deprecated models and scenarios): PASS — minimaxm31flash agentic-coding is active in MODELS.md as of 2026-11-10.

✅ Check 8 (Architecture hacks and all-reduce): PASS — There are no architecture overrides; sparse_disable_index_value matches the checkpoint's native config. TRTLLM_M31_AR_NVFP4 is a floating-point (E2M1) all-reduce, not an integer one. No ROCm quick-reduce controls apply. The TP2 eval lanes cover the NVFP4 all-reduce, the low-precision MoE combine and the NVFP4 KV.

✅ Check 9 (Spec-decode chat template): PASS — AgentX traffic goes through the Dynamo chat frontend, with DYN_CUSTOM_JINJA_TEMPLATE=/model/chat_template.jinja on every role.

✅ Check 10 (Engine patches): PASS — There are no patches, setup scripts or source edits. The only install is the published Dynamo wheel 1.7.0.dev20261028, and the TRTLLM_M31_* settings are runtime switches in the image.

✅ Check 11 (Golden AL): PASS — The c22 job log shows TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS=3.04. signoff_facts golden-al (dspark K7, thinking_on, standard) returns golden AL 4.04 from minimax-m3.1-flash__dspark__native__probabilistic-block.yaml, and 3.04 + 1 = 4.04.

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

✅ Check 14 (Pareto coverage): PASS — The single AgentX curve (minimaxm31flash, agentic-coding, gb300_dynamo-trt, fp4, P90 E2EL vs tput/GPU, run 42086507846 on 2f8becf2) has 7 of 10 measured points on the frontier per pareto_coverage, using the bmk_agentic_* artifacts. The dominated points are 1P2D TP4 c36 (dominated by 1P3D TP2 c36, not 1P2D TP2 c32 as the sign-off says) and 1P4D TP4 c36 and c40.

✅ Check 15 (PR description): PASS — The body's checked claims (10 benchmark and 9 eval jobs, GSM8K 96.0–97.0%, Dynamo 1.7.0.dev20261028, valid power on all points, AL setting 3.04) match the pinned run and recipes. The open TP2 draft-precision question is covered under Check 13.

➖ Check 16 (NVIDIA overview): N/A — Only TensorRT-LLM (dynamo-trt) configurations are affected, so there is no NVIDIA vLLM or SGLang pair.

✅ Check 17 (Metrics publication): PASS — Dynamo 1.7.0.dev20261028 is newer than ai-dynamo/dynamo#12162. Every role sets publish_metrics: true, enable_iter_perf_stats: true and iteration_stats_interval: 20. The artifacts report server_metrics.present: true with the trtllm adapter and KV-usage readings.

Assessed commit: 2f8becf2eb03e38c0a71091103372b2f76126200.

@SemiAnalysisAI SemiAnalysisAI deleted a comment from github-actions Bot Nov 10, 2026
allreduce_strategy: AUTO
moe_config:
backend: TRTLLM
use_low_precision_moe_combine: true

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

..

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

non-canary-full-sweep-enabled Run the full sweep without the canary gate (full search space, no trim)

Projects

Status: No status

Development

Successfully merging this pull request may close these issues.

6 participants