Skip to content

[AgentX] MiniMax-M3.1-Flash GB300 Dynamo TRT: add disaggregated points / [AgentX] MiniMax-M3.1-Flash GB300 Dynamo TRT:新增分离式运行点 - #5066

Open
Ankur-singh wants to merge 2 commits into
mainfrom
minimaxm31flash-fp4-gb300-dynamo-trt-agentic-dspark-disagg
Open

Ankur-singh wants to merge 2 commits into
mainfrom
minimaxm31flash-fp4-gb300-dynamo-trt-agentic-dspark-disagg

Conversation

@Ankur-singh

Copy link
Copy Markdown
Collaborator

Summary

Add seven disaggregated AgentX points for MiniMax-M3.1-Flash NVFP4 on GB300 with Dynamo and TensorRT-LLM (minimaxm31flash-fp4-gb300-dynamo-trt-agentic-dspark-disagg), using the feat-m31-flash-perf2-gb-clean-1.4.0a0.dev202611061956 TensorRT-LLM image and Dynamo 1.7.0.dev20261028. The points extend the TRT disaggregated curve from TP4 low-latency layouts up to DEP2 prefill with DEP8/DEP16 decode:

Layout Prefill Decode Concurrency
1P3D TP4 3 × TP4 36, 56, 76
1P4D TP4 4 × TP8 60
4P1D 4 × DEP2 DEP8 320
5P1D 5 × DEP2 DEP16 300
7P1D 7 × DEP2 DEP8 480

All points use DSpark draft length 7 with golden-curve synthetic acceptance, NVFP4 KV cache, native DRAM KV offload and NIXL transfer; the indexer KV cache is FP8 at 1P3D c36 and NVFP4 elsewhere. Every TensorRT-LLM worker sets stream_interval: 20, iteration and KV-cache statistics at interval 20, and publishes metrics. GPU and host CPU power collection are enabled on every point.

Overlap with #4751: #4751 adds the same …-dspark-disagg config key and disagg-variants.yaml on the earlier feat-m31-flash-perf1 image. This PR replaces that disaggregated half; #4751 should keep only its …-dspark-agg key before either PR merges, otherwise the two conflict.

Risks:

  • The image is a MiniMax-M3.1-Flash model-enablement build, not a main-branch TensorRT-LLM release. MODELS.md allows this temporarily for new models (MiniMax-M3.1-Flash was added 2026-10-23); these points need to move to a main-branch image once one carries the M3.1 support.
  • The DEP2 prefill workers set TRTLLM_M31_SM103_TP2_TARGET=1. Effective DSpark draft precision on that path still has to be confirmed from the sweep logs.
AI model disclosure
  • Model/version: claude-opus-5-5 and claude-sonnet-5-5 (Claude Code).
  • Role: claude-sonnet-5-5 did the initial config inspection; claude-opus-5-5 ported the config bundle, enabled power collection, wrote the changelog entry, ran the local checks and drafted this description. No delegated agents contributed.
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, a maintainer with write access to the repository 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

为 GB300 上的 MiniMax-M3.1-Flash NVFP4 新增七个 Dynamo + TensorRT-LLM 分离式 AgentX 运行点(minimaxm31flash-fp4-gb300-dynamo-trt-agentic-dspark-disagg),使用 TensorRT-LLM 镜像 feat-m31-flash-perf2-gb-clean-1.4.0a0.dev202611061956 和 Dynamo 1.7.0.dev20261028。覆盖范围从 TP4 低延迟布局延伸到 DEP2 prefill 搭配 DEP8/DEP16 decode:1P3D(TP4 / 3 × TP4)并发 36、56、76;1P4D(TP4 / 4 × TP8)并发 60;4P1D(4 × DEP2 / DEP8)并发 320;5P1D(5 × DEP2 / DEP16)并发 300;7P1D(7 × DEP2 / DEP8)并发 480。

所有运行点均使用 DSpark 草稿长度 7 及 golden-curve synthetic acceptance、NVFP4 KV cache、原生 DRAM KV offload 和 NIXL 传输;indexer KV cache 在 1P3D c36 为 FP8,其余运行点为 NVFP4。每个 TensorRT-LLM worker 设置 stream_interval: 20,迭代统计与 KV cache 统计间隔为 20,并发布 metrics。所有运行点均开启 GPU 及主机 CPU 功耗采集。

与 #4751 的重叠: #4751 基于较早的 feat-m31-flash-perf1 镜像,新增了相同的 …-dspark-disagg config key 和 disagg-variants.yaml。本 PR 取代其分离式部分;#4751 应在任一 PR 合并前只保留 …-dspark-agg key,否则两者会冲突。

风险:

  • 该镜像是 MiniMax-M3.1-Flash 的模型适配分支构建,而非 TensorRT-LLM 主分支发布版本。MODELS.md 允许新模型临时使用此类镜像(MiniMax-M3.1-Flash 于 2026-10-23 加入);待主分支镜像包含 M3.1 支持后,这些运行点需要迁移。
  • DEP2 prefill worker 设置了 TRTLLM_M31_SM103_TP2_TARGET=1,该路径下 DSpark 草稿模型的实际精度仍需通过 sweep 日志确认。

AI 模型使用说明:claude-sonnet-5-5(Claude Code)完成初步配置检查;claude-opus-5-5(Claude Code)移植配置、开启功耗采集、编写 changelog 条目、运行本地检查并起草本说明。没有委派其他 agent。

变更类型:配置变更。检查清单仅在上方填写。

…ints

Add seven NVFP4 disaggregated AgentX points for
minimaxm31flash-fp4-gb300-dynamo-trt-agentic-dspark-disagg on the
TensorRT-LLM feat-m31-flash-perf2-gb-clean image with Dynamo
1.7.0.dev20261028: TP4/TP8 1P3D and 1P4D points plus DEP2 prefill with
DEP8/DEP16 decode. GPU and host CPU power collection are enabled.

Co-authored-by: Pietro Cicotti <5833013+pcicotti@users.noreply.github.com>
@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。

@Ankur-singh Ankur-singh added the full-sweep-enabled Full sweep with canary gate; matrix jobs run to completion despite failures label Nov 13, 2026

@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.

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

Comment on lines +79 to +155
speculative_model: /model/dspark
cache_transceiver_config:
backend: NIXL
transceiver_runtime: PYTHON
max_tokens_in_buffer: 524288
kv_transfer_timeout_ms: 600000
kv_cache_bounce_size_mb: 1024
agent_bounce_buffer_enable: true
agent_bounce_params:
max_average_descriptor_size: 4MB
min_descriptor_count: '1'
enable_autotuner: true
enable_tokenization_cache: true
stream_interval: 20
reasoning_parser: minimax_m3
print_iter_log: false
enable_iter_perf_stats: true
iter_perf_stats_interval: 20
return_perf_metrics: true
backend: pytorch
gpus_per_node: 4
env_overrides:
TRTLLM_DISAGG_ROLE: context
decode:
env:
TLLM_LOG_LEVEL: INFO
TRTLLM_SERVER_DISABLE_GC: '1'
TRTLLM_WORKER_DISABLE_GC: '1'
TRTLLM_ENABLE_PDL: '1'
NCCL_GRAPH_MIXING_SUPPORT: '0'
PYTHONNOUSERSITE: '1'
TQDM_DISABLE: '1'
HF_HUB_OFFLINE: '1'
TRANSFORMERS_OFFLINE: '1'
PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
TRTLLM_KV_LAYERWISE_ONBOARD: '0'
TRTLLM_KV_ONBOARD_CTAS: '12'
TRTLLM_M31_NO_MIX: '1'
TRTLLM_M31_NO_MIX_DECODE_INTERVAL: '1'
TRTLLM_KV_TRANSFER_NUM_THREADS: '4'
PYTHONDONTWRITEBYTECODE: '1'
DYN_ROUTER_MODE: least-loaded
DYN_REASONING_PARSER: minimax_m3
DYN_TOOL_CALL_PARSER: minimax_m3
DYN_CUSTOM_JINJA_TEMPLATE: /model/chat_template.jinja
args:
max_num_tokens: 16384
disable_overlap_scheduler: false
cuda_graph_config:
enable_padding: true
trust_remote_code: true
max_seq_len: 1048576
attn_backend: TRTLLM
enable_chunked_prefill: true
allreduce_strategy: AUTO
moe_config:
backend: TRTLLM
use_low_precision_moe_combine: false
sparse_attention_config:
algorithm: minimax_m3
implementation: msa
sparse_disable_index_value: true
fuse_qkv_index_projection: true
kv_cache_config:
dtype: nvfp4
tokens_per_block: 128
use_kv_cache_manager_v2: true
free_gpu_memory_fraction: 0.94
iteration_stats_interval: 20
speculative_config:
decoding_type: DSpark
max_draft_len: 7
block_size: 7
use_rejection_sampling: true
advanced_sampling_mode: no_topk
attention_backend: TRTLLM
speculative_model: /model/dspark

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) All seven new points will publish their draft model as the container path /model/dspark. Every other MiniMax-M3.1-Flash DSpark result, and the golden curve, records native. TensorRT-LLM's speculative_model: /model/dspark reaches speculation_facts (infx/srt_slurm/spec_identity.py:48-54), which maps only Hugging Face cache paths to a repo and returns any other string unchanged. Fix: make speculation_facts report native for a draft loaded from inside the served checkpoint mount (/model/...). It must still map HF-cache paths to org/name, because the musespark TRT recipe relies on that.

Why this was flagged

For each row, bind_multinode calls bind_speculation (infx/srt_slurm/workload.py:479-496). That reads the decode role's speculative_config through spec_parameters (infx/srt_slurm/synthetic_acceptance.py:125-132), which returns model: /model/dspark. speculation_facts (spec_identity.py:51-54) checks only HF_CACHE_REPO, so it returns /model/dspark, and the binder writes SPEC_DRAFT_MODEL=/model/dspark into benchmark.env. speculation_fields (infx/results/metadata.py:45-48) then publishes it as spec_draft_model in every agentic result. The draft is the checkpoint's bundled DSpark head, and the docstring says such a draft should be reported as native. So these rows are the first to publish a container-local path in that field. Any comparison that groups MiniMax DSpark results by draft model will treat them as a different draft. The fingerprint excludes SPEC_DRAFT_MODEL (infx/matrix/fingerprint.py:22), and compare ignores draft, so no check catches this.

Verification: Both roles set speculative_model: /model/dspark (disagg-variants.yaml:79 and :155). speculation_facts (infx/srt_slurm/spec_identity.py:48-54) rewrites only paths that match HF_CACHE_REPO and otherwise returns model or "native", so the string /model/dspark passes through unchanged.

@Oseltamivir

Copy link
Copy Markdown
Collaborator
Screenshot 2026-11-13 at 10 53 50 PM Screenshot 2026-11-13 at 10 54 05 PM

@Ankur-singh TTFT/E2E needs improving

@Ankur-singh

Copy link
Copy Markdown
Collaborator Author

@Oseltamivir whats wrong? I don't see any problem with either TTFT/E2EL. Can you please clarify?

@github-actions

Copy link
Copy Markdown
Contributor

@Ankur-singh

Copy link
Copy Markdown
Collaborator Author

/use 43388252410

[by Claude Code]

@Ankur-singh

Ankur-singh commented Nov 18, 2026 •

Copy link
Copy Markdown
Collaborator Author

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 official upstream engine images built from main or master (or the equivalent default development branch), or meet the community-engine branch exception: the upstream maintainers (Inferact for vLLM, RadixArk for SGLang, and LightSeek for TokenSpeed) approve the branch/image for the submission, the upstream PRs must merge in a timely manner after submission, and the submission must then move to a main-branch image. For this exception, linked the upstream PRs, maintainer approval, and image build/source provenance in the additional detail section. New models using model-enablement branches must transition 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:

  • Assessed head: 28b7c85e36bf3d145dc6ab0402c00c2a2dd13b9c. New config minimaxm31flash-fp4-gb300-dynamo-trt-agentic-dspark-disagg (Dynamo + TensorRT-LLM on GB300): 1P3D TP4/3xTP4 c36/c56/c76, 1P4D TP4/4xTP8 c60, 4P1D 4xDEP2/DEP8 c320, 5P1D 5xDEP2/DEP16 c300, 7P1D 7xDEP2/DEP8 c480; DSpark draft length 7, NVFP4 KV, NIXL.
  • PR validation + evals: run 43388252410 on the assessed head (attempt 4, success): 7/7 benchmarks and 7/7 gsm8k evals succeed, no later commits. Attempt 1's c480 benchmark failed on a pyxis/enroot unsquashfs error (infra), attempt 2 was cancelled, attempt 3 reran c480 and the evals, and attempt 4 reran only the c60 eval after a missing-Dynamo-wheel error on one runner (infra). /use 43388252410 is posted.
  • Evals: 13 of 14 eval rows (two per point, attempts 1 and 3/4) are 0.9568–0.9644 em_strict, within threshold of each other and of the MiniMax-M3.1-Flash TRT cohort. The c480 attempt-3 row is 0.9469, 0.0174 below the best sibling (c76 0.9644, threshold 0.0160) but within threshold of its cohort (VR200 c200 0.9613); the c480 attempt-1 eval on the same SHA and recipe scored 0.9606, so I read the attempt-3 row as run-to-run variance rather than an accuracy loss on the 7P1D path. 2(c): the c36 eval runs the FP8 indexer cache (indexer_kv_dtype: fp8 in its effective configs), the other six the NVFP4 one; the 4P1D eval's prefill logs show TRTLLM_M31_SM103_TP2_TARGET=1 engaging ModelingV2MiniMaxM31Nvfp4Sm103Dep2Ctx on all four DEP2 prefill workers.
  • Golden AL: every benchmark job sets TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS=3.04 on prefill and decode, i.e. AL 4.04 at 7 draft tokens from minimax-m3.1-flash__dspark__native__probabilistic-block (golden-al --framework dynamo-trt --pinned 3.04 matches). The report-only AL gate on real acceptance measured 3.84–4.01 on the TP4 points and 3.81–3.83 (0.943–0.949x) on the DEP points.
  • Draft precision: the shipped dspark/ head (MXFP8 q/k/v/o and MLP linears; BF16 norms and fc/markov/confidence heads) loads from /model/dspark with no draft quant, dtype or KV setting in the recipe. Pinned image (feat-m31-flash-perf2-gb-clean-1.4.0a0.dev202611061956, arm64 sha256:0d8a56b1…, TensorRT-LLM ce15082f) source: the drafter config drops quantization_config (_torch/pyexecutor/config_utils.py:976-990) and prepare_minimax_dspark_weights dequantizes MXFP8 to BF16 losslessly (_torch/models/modeling_dspark.py:3563-3593), with no requantization. TP4 points (c36/c56/c76, 1P4D c60) use M3DSparkWorker, where the image-default MXDRAFT recovers the bit-exact MXFP8 weights and runs the decode GEMVs in MXFP8 (logs: M3.1 DSpark MXDRAFT: MXFP8 drafter GEMV on 20 projections, i.e. all 5 layers x 4). DEP points (c300/c320/c480) run the same weights in BF16. Draft KV is BF16 on both paths (TP4: draft KV cache manager pool: 5 x (…) torch.bfloat16; DEP: External drafter KV pool keeps the drafter dtype; dropping the KV quantization inherited from the target.). TRTLLM_M31_SM103_TP2_TARGET does not reach the draft (dep2ctx route; no M31 h84 lines). The draft-touching env flags in the recipe only fuse launches or reorder sums, per the pinned source (_torch/_experimental/modeling_v2/models/minimax_m3/…): TRTLLM_M31_ENABLE_DRAFT_FAST (c60 decode; …tp8/dspark.py:117-125: "same math and precision; the summation order … changes, so not bitwise", its split-KV attention keeps bf16 Q/K/V with fp32 logits/softmax, …tp8/ops/attention/m3_draft_swa_decode.py), TRTLLM_M31_ENABLE_DRAFT_PREP_FUSE (c60; …tp8/dspark.py:354-358, …tp8/ops/gemm/m31_fc_shard.py:10-18: the fc gather as bf16 words, "the same bits for every non-NaN value"), TRTLLM_M31_ENABLE_DRAFT_LSE_FOLD (c60; _torch/speculative/dflash_lse_fold.py:3-20: same LSE instructions, each shape folds only after a bitwise-equal check against the stock path) and TRTLLM_M31_ENABLE_DRAFT_PREP_FUSE_TP4 (c56/c76; …tp4/dspark.py:81-89: ctx advance + read view as one copy launch). None changes a draft dtype or quantization. No submitter override, substitution or quantization.
  • Image: nvcr.io/nvidia/tensorrt-llm/release:feat-m31-flash-perf2-gb-clean-1.4.0a0.dev202611061956, NVIDIA's official TensorRT-LLM nvcr org, resolves (tag, not digest). It is a MiniMax-M3.1-Flash enablement build (the public NVIDIA/TensorRT-LLM feat/minimax-m3.1-flash line); MiniMax-M3.1-Flash is a new model (MODELS.md, 2026-10-23), so this uses the new-model transition allowance, and the config should move to a main-branch TensorRT-LLM release once one carries M3.1 support. Dynamo 1.7.0.dev20261028 wheel. Engine-first ordering: GB300 already has vLLM and SGLang (plus Dynamo) MiniMax-M3.1-Flash entries on main. No setup script or engine patch.
  • Recipe / upstream-image items: N/A, Dynamo + TensorRT-LLM only (no vLLM/SGLang serving config). Not append-only.
  • Dynamo TRT-LLM metrics: workers run with --publish-metrics, effective configs set enable_iter_perf_stats: true, iteration/KV stats at interval 20 and return_perf_metrics; every bmk result has server_metrics.present with the trtllm adapter (213–215 metrics). stream_interval: 20 on every worker. No hf-overrides; sparse block settings match the HF config; allreduce_strategy: AUTO, no integer all-reduce.
  • Pareto (run 43388252410, total tput/GPU vs p90 E2EL, curve gb300_dynamo-trt / minimaxm31flash / agentic / fp4): 7 of 7 points on the frontier.
  • Team TTFT gate: p90 TTFT c76 4.72, c60 4.44, c300 1.90, c320 2.25, c480 2.19 s; c36 (5.96 s) and c56 (5.70 s) are in the 5–10 s fallback band: they are the 1P3D low-latency points with a single TP4 prefill worker, where prefill queueing on bursty agentic turns sets the tail (p50 0.41/0.84 s). No point reaches 10 s.
  • Known follow-ups: Add MiniMax M3.1 Flash GB300 Dynamo TRT AgentX configurations / 新增 MiniMax M3.1 Flash GB300 Dynamo TRT AgentX 配置 #4751 (open) still adds the same -dspark-disagg key and disagg-variants.yaml on the earlier perf1 image and must keep only its -dspark-agg key (or close) before either merges. The PR conflicts with main in perf-changelog.yaml only. The body's "native DRAM KV offload" applies to the TP4 and DEP prefill workers; DEP decode runs with host_cache_size: 0. Results record the draft as /model/dspark rather than native; that is a result-identity issue in infx/srt_slurm/spec_identity.py, not this config.

Signed: Ankur-singh

[by Claude Code]

@github-actions

github-actions Bot commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

⚠️ Verdict: WARN ⚠️

⚠️ Check 13 (Draft precision): WARN — Draft precision could not be verified for minimaxm31flash-fp4-gb300-dynamo-trt-agentic-dspark-disagg. The recipe sets no draft quantization, dtype or KV override and loads the shipped /model/dspark head, but the sign-off's evidence rests on source in the pinned image's TensorRT-LLM commit ce15082f, which is not public (NVIDIA/TensorRT-LLM returns "No commit found"). The public feat/minimax-m3.1-flash tip 7fc3d05c has none of the cited files (…/minimax_m3/tp8/dspark.py, …/tp4/dspark.py, m3_draft_swa_decode.py, m31_fc_shard.py, dflash_lse_fold.py). The cited log lines (MXDRAFT: MXFP8 drafter GEMV on 20 projections, BF16 draft KV pool) are only inside multinode_server_logs tarballs of run 43388252410, which this verifier could not extract. The GitHub job logs show the draft-touching flags being set (TRTLLM_M31_ENABLE_DRAFT_FAST, _DRAFT_LSE_FOLD, _DRAFT_PREP_FUSE on c60 decode; _DRAFT_PREP_FUSE_TP4 on c56/c76 decode) but not their effect. Human review needed: confirm from the image source and server logs that MXDRAFT keeps MXFP8 weights, the DEP path runs BF16, draft KV stays BF16, and these flags only fuse or reorder work.

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 **/gb300-*/ srt-slurm recipe path. perf-changelog.yaml falls only under the catch-all owner.

✅ Check 1 (Sweep on in-PR commit): PASS — The PR head is still the pinned SHA 28b7c85e. Run 43388252410 (attempt 4, success, head_sha 28b7c85e) has 7/7 executed multi-node agentic/canary benchmark checks and 7/7 multi-node agentic eval checks green. There are no later commits.

✅ Check 2 (Evals): PASS — gsm8k ran on all 7 points with the PR image feat-m31-flash-perf2-gb-clean-1.4.0a0.dev202611061956, scoring 0.9469–0.9644 em_strict (c56 artifact confirms 0.9636 ± 0.0052). Every row is within threshold of its cohort. Only the c480 attempt-3 row (0.9469) misses the sibling bar against c76 (0.9644; gap 0.0174, threshold 0.0160). The c480 attempt-1 eval on the same SHA and recipe scored 0.9606 (ok), and the sign-off records that rerun. Each recipe variant, including the c36 FP8 indexer and the DEP2 TRTLLM_M31_SM103_TP2_TARGET prefill, has its own eval.

➖ Check 3 (Upstream recipe): N/A — The only affected config is framework: dynamo-trt (Dynamo + TensorRT-LLM), so no vLLM or SGLang serving configuration is affected.

✅ Check 4 (Reuse command): PASS — Ankur-singh (permission write, role write) posted /use 43388252410 on its own line, which is the Check 1 run.

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

✅ Check 6 (Upstream images / ordering): PASS — (a) does not apply because there is no vLLM/SGLang entry. (b) GB300 already has minimaxm31flash vLLM, SGLang, dynamo-sglang and dynamo-vllm entries in nvidia-master.yaml. (c) The image is NVIDIA's official nvcr TensorRT-LLM build from the feat/minimax-m3.1-flash enablement line. MiniMax-M3.1-Flash is a new model (added 2026-10-23), so the temporary allowance applies, and the sign-off records the plan to move to a main-branch image.

✅ Check 7 (Deprecated models): PASS — minimaxm31flash agentic coding is active in MODELS.md as of 2026-11-19.

✅ Check 8 (No arch hacks): PASS — There are no hf-overrides or model-file edits. sparse_init_blocks: 0 and sparse_local_blocks: 1 are sparse-attention settings, not layer or expert skipping. allreduce_strategy: AUTO on NVIDIA, with no integer all-reduce controls set, and low-precision MoE combine is off. The FP8/NVFP4 indexer KV choices are covered by per-point evals.

✅ Check 9 (Chat template): PASS — The agentic replay goes through the Dynamo chat endpoint, with DYN_CUSTOM_JINJA_TEMPLATE: /model/chat_template.jinja and the minimax_m3 reasoning and tool parsers on both roles.

✅ Check 10 (No engine patches): PASS — The recipe and master entry have no patch, setup script, source edit or engine-wheel override. Only the published Dynamo 1.7.0.dev20261028 wheel is installed, through dynamo.install.

✅ Check 11 (Golden AL): PASS — Job logs show TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS=3.04 applied to the prefill and decode roles. That is golden AL 4.04 for 7 draft tokens from minimax-m3.1-flash__dspark__native__probabilistic-block.yaml (golden-al reports matches: true).

➖ Check 12 (Append-only): N/A — The new changelog entry does not set append-only: true.

✅ Check 14 (Pareto coverage): PASS — Curve minimaxm31flash / agentic-coding / gb300_dynamo-trt / fp4 / p90 E2EL, from run 43388252410 at SHA 28b7c85e (bmk_agentic_* artifacts): 7/7 points are on the frontier. Throughput per GPU rises monotonically with E2EL, from 14334 at 6.02 s to 134211 at 14.53 s. The helper's stdin input was blocked by the sandbox, so this count was computed by hand.

✅ Check 15 (PR description): PASS — The title, body and changelog match the recipe on topology, concurrency, image, Dynamo version, NVFP4 KV, the FP8 indexer at c36, stream_interval 20, and TRTLLM_M31_SM103_TP2_TARGET on DEP2 prefill.

➖ Check 16 (NVIDIA overview): N/A — The affected config is TensorRT-LLM only, so no vLLM/SGLang model/framework pair applies.

✅ Check 17 (Metrics publication): PASS — Dynamo 1.7.0.dev20261028 postdates ai-dynamo/dynamo#12162 (merged 2026-09-03). Job logs show every prefill and decode worker launching with --publish-metrics, and enable_iter_perf_stats: true is set with iteration and KV stats at interval 20.

✅ Check 18 (Recipe consistency): PASS — All 7 keys generate. Master worker counts, TP/EP/DP-attn and GPU/node math match each recipe override. The image resolves, and Dynamo is pinned to a published wheel. No points were removed or retuned, and no regressions were found. stream_interval: 20 (at most 60) is set on every prefill and decode worker.

Assessed commit: 28b7c85e36bf3d145dc6ab0402c00c2a2dd13b9c.

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

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

Projects

Status: No status

Development

Successfully merging this pull request may close these issues.

2 participants