[Klaud Cold] minimaxm3-fp8-h100-vllm-agentic-mtp: MiniMax-M3 H100 AgentX EAGLE3 recipe with golden-AL synthetic acceptance / MiniMax-M3 H100 AgentX EAGLE3 配方（黄金 AL 合成接受）

[functionstackx]

Summary

Adds minimaxm3-fp8-h100-vllm-agentic-mtp — the EAGLE3 speculative-decoding (spec-decoding: mtp) variant of the MiniMax-M3 MXFP8 H100 AgentX recipe. One new benchmark script, one master-config entry, one changelog entry; no launcher change is needed (launch_h100-dgxc-slurm.sh already composes agentic/ + _mtp).

Speculative config — method: eagle3 on Inferact/MiniMax-M3-EAGLE3, num_speculative_tokens: 3, drafter pinned to FLASH_ATTN (the EAGLE3 head is MHA; FlashInfer only serves the mandatory page size 128 via its trtllm-gen kernel, which needs GQA/MQA). Same draft and level as every merged MiniMax-M3 MTP recipe.

Synthetic forced AL — throughput runs set rejection_sample_method: synthetic with synthetic_acceptance_length: 2.83, the committed golden AL for MiniMax-M3 EAGLE3 at K=3, thinking_on (golden_al_distribution/minimaxm3_eagle3.yaml). Per the AgentX policy in golden_al_distribution/README.md, a submission may choose any supported draft length but may not substitute a different acceptance target. The separate minimaxm3_eagle3_gqa.yaml curve belongs to the Inferact/MiniMax-M3-EAGLE3-GQA draft (run 29784780049) and is not mixed in. EVAL_ONLY runs drop synthetic acceptance and keep real verification, since synthetic acceptance commits drafted tokens regardless of the target logits and would score the eval at ~0.

Serve shape is otherwise identical to the non-MTP agentic sibling (Mooncake host-DRAM KV offload, --block-size 128, --language-model-only, --kv-cache-dtype fp8, --attention-backend TRITON_ATTN, minimax_m3 parsers, gmu 0.90, --cpu-offload-gb 26, vllm-router for DP-attention) so the spec-decode delta is readable at equal concurrency. max-num-seqs stays 2*CONC for AgentX subagent fan-out; cudagraph capture is raised to MAX_NUM_SEQS*(1+3) tokens because vLLM's FULL_DECODE_ONLY ladder is token-denominated.

Search space covers four arms (TP8 and TP8/EP8 × GPU-resident and Mooncake) on the AgentX MTP concurrency grid: GPU-resident rows [1, 4, 8, 12, 16], KV-offload rows [4, 8, 12, 16]. Every step is at least 2 concurrency apart and every arm stops hard at conc 16 — single-step sampling cannot separate configurations by more than run-to-run noise on the agentic corpus, and past 16 these SKUs are into the post-HBM-cliff thrashing regime the non-MTP sweeps already characterized. 18 generated jobs.

Validation: bash -n clean; both YAMLs parse; generate_sweep_configs.py test-config emits 18 agentic-coding configs with spec-decoding: mtp and the expected TP/EP/KV arms.

中文说明

新增 minimaxm3-fp8-h100-vllm-agentic-mtp：MiniMax-M3 MXFP8 H100 AgentX（智能体编码）配方的 EAGLE3 投机解码（spec-decoding: mtp）变体。包含一个新的基准测试脚本、一条 master 配置条目和一条 changelog 条目；启动器无需改动（launch_h100-dgxc-slurm.sh 已能拼出 agentic/ + _mtp 路径）。

投机解码配置：method: eagle3，草稿头 Inferact/MiniMax-M3-EAGLE3，num_speculative_tokens: 3，drafter 固定 FLASH_ATTN（EAGLE3 头是 MHA，而 FlashInfer 仅通过要求 GQA/MQA 的 trtllm-gen kernel 支持强制的 page size 128）。与仓库中所有已合并的 MiniMax-M3 MTP 配方使用相同的草稿头与投机层数。

合成强制 AL：吞吐量测试设置 rejection_sample_method: synthetic 与 synthetic_acceptance_length: 2.83，即 MiniMax-M3 EAGLE3 在 K=3、thinking_on 下已提交的黄金 AL（golden_al_distribution/minimaxm3_eagle3.yaml）。依据 golden_al_distribution/README.md 的 AgentX 规则，提交方可自选草稿长度，但不得替换接受长度目标。minimaxm3_eagle3_gqa.yaml 对应的是 Inferact/MiniMax-M3-EAGLE3-GQA 草稿头（run 29784780049），两条曲线不混用。EVAL_ONLY 场景关闭合成接受、保留真实验证，否则生成文本错误、评估分数接近 0。

服务参数与非 MTP 版本完全一致（Mooncake 主机 DRAM KV 卸载、--block-size 128、--language-model-only、--kv-cache-dtype fp8、--attention-backend TRITON_ATTN、minimax_m3 解析器、gmu 0.90、--cpu-offload-gb 26、DP-attention 下的 vllm-router），便于在相同并发下直接对比投机解码带来的差异。max-num-seqs 保持 2*CONC 以容纳 AgentX 子智能体扇出；cudagraph 捕获上限提高到 MAX_NUM_SEQS*(1+3) 个 token，因为 vLLM 的 FULL_DECODE_ONLY 阶梯是按 token 计的。

搜索空间覆盖四条分支（TP8 与 TP8/EP8 × GPU 常驻与 Mooncake），并采用 AgentX MTP 并发网格：GPU 常驻 [1, 4, 8, 12, 16]，KV 卸载 [4, 8, 12, 16]。相邻并发间隔至少为 2，且一律止于并发 16 —— 在智能体语料上，逐 1 采样无法区分出超过运行间噪声的差异；而并发超过 16 后，这些 SKU 已进入非 MTP 扫描早已刻画过的显存瓶颈后抖动区间。共生成 18 个作业。

验证：bash -n 通过；两个 YAML 均可解析；generate_sweep_configs.py test-config 生成 18 条 agentic-coding 配置，spec-decoding: mtp 与 TP/EP/KV 分支均符合预期。

🤖 Generated with Claude Code

---

Revision (post-first-sweep)

1. --gpu-memory-utilization 0.90. The first agentic job on [Klaud Cold] minimaxm3-fp8-h100-vllm-agentic-mtp: MiniMax-M3 H100 AgentX EAGLE3 recipe with golden-AL synthetic acceptance / MiniMax-M3 H100 AgentX EAGLE3 配方（黄金 AL 合成接受） #2423 (TP8/EP8 c9, run 30515793863) died mid-warmup: all 8 ranks raised torch.OutOfMemoryError 96 MiB short, killing the engine core (EngineDeadError) and tripping AIPerf's terminal warmup failure. The recipe had inherited the non-MTP agentic sibling's gmu, which has no draft head to house; the proven MTP sibling on each of these SKUs runs 0.90 (explicit on H100/H200, vLLM's default on the ROCm ones). Applied to all four MiniMax PRs.
2. Concurrency grid. Steps of at least 2 and a hard stop at conc 16, replacing the dense integer sampling copied from the STP siblings.

中文：(1) --gpu-memory-utilization 改为 0.90 —— #2423 的首个智能体作业（TP8/EP8 c9，run 30515793863）在预热阶段 8 个 rank 全部 OOM（仅差 96 MiB），导致 engine core 崩溃并触发 AIPerf 终止性预热失败；此前沿用了非 MTP 版本的取值，而该版本无需为草稿头预留显存，各 SKU 上已验证的 MTP 配置均为 0.90。四个 MiniMax PR 均已修正。(2) 并发网格：相邻间隔至少为 2，且一律止于 16。

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档

[functionstackx]

Draft-head note for reviewers. This recipe uses Inferact/MiniMax-M3-EAGLE3 with golden AL 2.83 (minimaxm3_eagle3.yaml, thinking_on, K=3), matching the fixed-seq-len MTP sibling on this same SKU.

Worth knowing: #2337 (merged 2026-07-29) moved B200/B300 fp8 to the Inferact/MiniMax-M3-EAGLE3-GQA head, with the drafter backend going TRITON_ATTN → FLASH_ATTN and K kept at 3. The Hopper and gfx942 fixed-seq-len MTP recipes were not part of that swap and still run the non-GQA head, so this PR follows them — draft and curve stay matched, which is the part that matters for the AgentX contract. If you want the Hopper/ROCm SKUs to follow Blackwell onto the GQA head, it is a two-line change here: the draft id, and AL 2.83 → 2.78 (minimaxm3_eagle3_gqa.yaml, thinking_on, K=3). Say the word and I will make that swap before the sweep spends much more GPU time.

中文：本配方使用 Inferact/MiniMax-M3-EAGLE3 草稿头，黄金 AL 为 2.83（minimaxm3_eagle3.yaml，thinking_on，K=3），与同 SKU 的固定序列长度 MTP 配方保持一致。

需要说明的是：#2337（2026-07-29 合并）已将 B200/B300 fp8 切换到 Inferact/MiniMax-M3-EAGLE3-GQA 草稿头，drafter 注意力后端由 TRITON_ATTN 改为 FLASH_ATTN，K 仍为 3。Hopper 与 gfx942 的固定序列长度 MTP 配方未包含在该次切换中，仍使用非 GQA 草稿头，因此本 PR 与之保持一致 —— 关键在于草稿头与 AL 曲线相互匹配，这正是 AgentX 规则所要求的。若希望 Hopper/ROCm SKU 也跟随 Blackwell 切到 GQA 草稿头，只需改两处：草稿头 ID，以及 AL 2.83 → 2.78（minimaxm3_eagle3_gqa.yaml，thinking_on，K=3）。如需切换请告知，我会在扫描消耗更多 GPU 前完成。

[claude]

🔴 Lines 175-176 pass both --tool-call-parser minimax_m3 and --reasoning-parser minimax_m3, but minimaxm3_fp4_mi355x.sh (merged the day before, PR #2118) intentionally omits the reasoning parser for this exact agentic-coding scenario, with a detailed comment explaining that MiniMax-M3's <mm:think> block must round-trip in message.content every turn or multi-turn quality collapses into repeated commands / empty patches. This is inherited from minimaxm3_fp8_h100.sh and its h200/mi300x/mi325x siblings, but this PR copies the same known-bad flag pair into brand-new code, and it hits hardest on the EVAL_ONLY real-verification path (lines 195-196), which would score near zero from the resulting degeneration. Fix: drop --reasoning-parser minimax_m3 here as was already done in the mi355x recipe.

Extended reasoning...

The bug: minimaxm3_fp8_h100_mtp.sh:175-176 passes both --tool-call-parser minimax_m3 and --reasoning-parser minimax_m3 to the vLLM server. MiniMax-M3 is an interleaved-thinking model that emits an <mm:think>...</mm:think> block as part of its response. When --reasoning-parser minimax_m3 is enabled, vLLM strips that block out of message.content and surfaces it instead as a separate, response-only reasoning_content field on the OpenAI-compatible completion object.

Why that breaks multi-turn agentic runs: the agentic-coding harness this script drives (run_agentic_replay_and_write_outputs, mini-swe-agent + litellm's OpenAI client) only round-trips message.content back into the conversation on the next turn — it does not resend reasoning_content. So with the reasoning parser enabled, the model's think block silently disappears from its own conversation history every turn. Without visibility into its prior plan, the model degenerates: it repeats the same command turn after turn until the step limit is hit, producing an empty patch.

Why nothing here prevents it: this exact failure mode and fix are already documented in the repo. benchmarks/single_node/agentic/minimaxm3_fp4_mi355x.sh:119-127 (PR #2118, merged one day before this recipe's non-MTP sibling) keeps --tool-call-parser minimax_m3 but explicitly omits --reasoning-parser minimax_m3, with an inline comment spelling out precisely this mechanism. That fix was never backported to minimaxm3_fp8_h100.sh (nor its h200/mi300x/mi325x siblings), and this new MTP recipe copies the non-MTP h100 sibling's server flags verbatim (including both parsers), propagating the known-bad configuration into new code rather than following the already-established fix.

Impact: the failure is model/harness-level, not hardware- or spec-decode-specific, so it applies identically on this H100 EAGLE3 recipe. The throughput sweep runs use rejection_sample_method: synthetic with a forced golden acceptance length, so the actual generated text there is synthetic/garbage by design and the parser bug doesn't change the throughput numbers. But the EVAL_ONLY path (lines 195-196) drops synthetic acceptance specifically to do real target verification and score genuine agentic quality — this is exactly the path that would collapse into empty-patch degeneration from the stripped think block, defeating the purpose of carving out EVAL_ONLY in the first place.

Step-by-step proof of the mechanism:

1. Server starts with --reasoning-parser minimax_m3 (line 176), so every completion response has its <mm:think>...</mm:think> content removed from message.content and placed only in message.reasoning_content.
2. Turn N: the agent (mini-swe-agent via litellm's OpenAI-compatible client) receives a response, appends it to its running message list. It appends message.content (the tool-call/action portion) — reasoning_content is not part of the OpenAI chat schema the client resends, so it's dropped.
3. Turn N+1: the server serves the next completion using only the truncated history — the model has no record of its own prior <mm:think> reasoning/plan.
4. Because MiniMax-M3 relies on that block persisting in-context to track multi-step plans (per the mi355x comment), the model re-derives a similar-looking next step, often repeating its previous command verbatim.
5. This repeats every turn until the step limit, producing no net progress and an empty final patch — which the EVAL_ONLY real-verification path would score at or near zero, even though the server and harness are otherwise functioning correctly.

Fix: drop --reasoning-parser minimax_m3 from the vllm serve invocation (keep --tool-call-parser minimax_m3), matching what minimaxm3_fp4_mi355x.sh already does for this identical scenario.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30515793863
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30515793863

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30517270909
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30517270909

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=30517356982
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=30517356982