perf(amd): enable MiniMax-M3 ATOM FlyDSL decode / 启用 MiniMax-M3 ATOM FlyDSL 解码

[yhl-amd]

Summary

Enable the MiniMax-M3 paged-decode optimization from ROCm/ATOM#2366 in the existing MI355X AgentX recipe, following #3189 and the upstream recipe.

• Update rocm/atom-dev:nightly_202609171455 to rocm/atom-dev:nightly_202609231248. Enable ATOM_PA_FLYDSL=1 and ATOM_PA_FLYDSL_PLAN=1; changing the image alone would leave FlyDSL disabled.
• Fix GPU visibility: when ROCR supplies physical GPU IDs, clear HIP's second mask. Automatically selected offload GPUs use ROCR alone (0,4 for TP2; 0,1,4,5 for TP4). Preserve caller-supplied ROCR and HIP-only allocations.
• Retain all 19 throughput points, 3600-second profiles, EAGLE3 K3, golden AL 2.78, indexer CP, graph sizes, and existing LMCache budgets. Eval keeps real draft acceptance. The existing Inferact/MiniMax-M3-EAGLE3-GQA draft and precision settings are unchanged; no engine patches are applied.

Registry manifest: sha256:80f2c66579704aba933a9ff2a6e22e1dc61917c0fe6a2b0a0166816cd6d0100a. Image labels identify ATOM 0.1.7.dev16+g10ab5e8c1, source revision 10ab5e8c1995c3e8d9e0674534f093d1e33a04ff, and ROCm 7.2.4. That ATOM revision contains #2366. Actual kernel routing and performance remain to be verified in GPU CI.

CI scope

One PR sweep contains all 19 throughput jobs and exactly one eval job: TP4 C48 with LMCache offload, using minimax-vendor / minimax_m3_full (102 provider cases) and real draft acceptance.

The changelog sets eval-concs: [48]. The planner filters only already-selected eval concurrencies, leaves throughput intact, and rejects a selection with no matching evals. The field also narrows multi-node all-concurrency eval rows and is incompatible with no-evals / append-only.

Combined GPU CI started: Run Sweep #35873197630, on commit 7bbb0ec775dd86b8f6ebb3c4e345cdfe283b1c1a, includes throughput and the single eval in one workflow. The two previous separate runs were cancelled.

Validation

• Bash syntax, YAML parsing, git diff --check, and the real changelog validator passed; historical changelog bytes are preserved.
• The real changelog planner emits 19 AgentX throughput jobs and exactly one TP4 C48 offload eval in the same plan; other eval buckets are empty.
• 450 focused changelog, matrix, and validation tests passed, including eval filtering, unchanged throughput, multi-node selection, invalid inputs, and incompatible modifiers. Ruff lint and format checks passed for all infx/ Python files.
• Ran the real script and shared replay builder with external GPU, network, installation, and workload operations stubbed: inherited ROCR/HIP masks, automatic TP2 and TP4 offload placement, explicit ROCR placement, and HIP-only placement passed. Verified FlyDSL/planner enablement, per-rank memory/lookup workers, TP-specific indexer CP, real-acceptance eval, and chat-template replay.
• GPU throughput and eval results are pending. No new measured performance claims are made.

AI model disclosure

Prepared with OpenAI GPT-6 via Codex for implementation, source inspection, and local validation. The exact runtime model/version identifier was not exposed and could not be verified. No delegated agents were used.

中文

概要

参考 #3189 和 ATOM 上游配方，将 ROCm/ATOM#2366 的 MiniMax-M3 paged-decode 优化接入现有 MI355X AgentX 配置。

• 镜像从 rocm/atom-dev:nightly_202609171455 更新为 rocm/atom-dev:nightly_202609231248，显式启用 ATOM_PA_FLYDSL=1 和 ATOM_PA_FLYDSL_PLAN=1，避免只换镜像却未打开新 kernel。
• 修复 GPU 可见性：ROCR 已提供物理 GPU ID 时清除 HIP 的第二层 mask。自动选择的 offload GPU 仅通过 ROCR 设置，TP2 为 0,4，TP4 为 0,1,4,5；保留调用方的 ROCR 分配及 HIP-only 分配。
• 保留全部 19 个吞吐点、3600 秒测量时长、EAGLE3 K3、golden AL 2.78、indexer CP、图捕获尺寸和现有 LMCache 预算。eval 使用真实 acceptance；现有 Inferact/MiniMax-M3-EAGLE3-GQA draft 及精度设置保持不变，不修改 engine。

Registry manifest 为 sha256:80f2c66579704aba933a9ff2a6e22e1dc61917c0fe6a2b0a0166816cd6d0100a。镜像标签记录 ATOM 0.1.7.dev16+g10ab5e8c1、源码 revision 10ab5e8c1995c3e8d9e0674534f093d1e33a04ff 和 ROCm 7.2.4，该 ATOM revision 已包含 #2366。实际 kernel 路由和性能仍待 GPU CI 确认。

CI 范围

同一次 PR sweep 包含全部 19 个吞吐 job 和唯一一个 eval job：TP4 C48、LMCache offload，使用 minimax-vendor / minimax_m3_full（102 个 provider case）和真实 draft acceptance。

changelog 设置 eval-concs: [48]。planner 只过滤常规策略已选中的 eval 并发，保留全部吞吐点；没有匹配 eval 时拒绝生成。该字段也支持缩减多节点的全并发 eval 行，不能与 no-evals / append-only 同时使用。

合并后的 GPU CI 已启动： Run Sweep #35873197630，运行提交为 7bbb0ec775dd86b8f6ebb3c4e345cdfe283b1c1a，在同一条工作流中包含吞吐和唯一的 eval。此前两条分开的运行已取消。

验证

• Bash 语法、YAML 解析、git diff --check 和正式 changelog validator 均通过，历史 changelog 字节保持不变。
• 正式 changelog planner 在同一份 plan 中生成 19 个 AgentX 吞吐 job 和唯一一个 TP4 C48 offload eval，其他 eval bucket 均为空。
• 450 项 changelog、矩阵和校验测试通过，覆盖 eval 过滤、保留吞吐、多节点选择、非法输入及冲突选项；全部 infx/ Python 文件通过 Ruff lint 和格式检查。
• 使用真实脚本和公共 replay builder，仅替代外部 GPU、网络、安装及工作负载操作，验证了继承 ROCR/HIP、自动 TP2/TP4 offload、显式 ROCR 和 HIP-only 分配，以及 FlyDSL/planner 开关、每 rank 内存与 lookup worker、TP 限定的 indexer CP、真实 acceptance eval 和 chat-template replay。
• GPU 吞吐和 eval 结果待完成，本 PR 不声称已获得新的实测性能结果。

AI 模型披露

由 OpenAI GPT-6 通过 Codex 完成实现、源码检查和本地验证。运行环境未暴露确切模型／版本标识，因此无法核实。未使用委派代理。

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps only run on labeled PRs. Add full-sweep-fail-fast (strongly recommended); use full-sweep-enabled only when matrix jobs should continue after a failure.
• 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：**扫描仅在带有标签的 PR 上运行。强烈建议添加 full-sweep-fail-fast；仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。
• **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。

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=36233199454

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=36233199454

[claude]

🔴 The new changelog entry ships pr-link: TBD even though this PR is already open, so the merge-prep tooling rejects it instead of finalizing the link automatically. validate_added_pr_link in infx/workflows/validate_perf_changelog.py only accepts "XXX" or the canonical https://github.com/SemiAnalysisAI/InferenceX/pull/ as placeholders (PR_LINK_PLACEHOLDERS, line 21-24); "TBD" matches neither. docs/configuration-procedures.md:584 says TBD is only for use before the PR exists and must be replaced right after creation, which this entry was not. Fix: replace pr-link: TBD with either the real https://github.com/SemiAnalysisAI/InferenceX/pull/ URL for this PR or the XXX placeholder that the validator/canonicalizer actually recognizes.

Why this was flagged

perf-changelog.yaml:8771 sets pr-link: TBD on the newly appended minimaxm3 entry. infx/workflows/prepare_perf_changelog_merge.py's canonicalize_appended_links (used to finalize pr-links for reuse-assisted merges) calls compare_entries -> validate_added_pr_link with the real pr_number; since "TBD" is not in PR_LINK_PLACEHOLDERS ({"XXX", ".../pull/XXX"}) and not the expected canonical link, it raises ChangelogValidationError("new PR entry must use ... or an XXX placeholder; found 'TBD'"), and canonicalize_appended_links's own check at prepare_perf_changelog_merge.py:106-109 raises "appended entry N has unexpected pr-link 'TBD'" too. On base branch, entries use the recognized "XXX" placeholder, so this step succeeds; here whoever runs the merge-prep/canonicalize step for this PR gets a hard failure instead of the link being auto-filled, blocking merge until manually fixed.

Verification: normal. The new entry at perf-changelog.yaml:8771 ships pr-link: TBD. This value is not an accepted placeholder: validate_perf_changelog.py:21-24 defines PR_LINK_PLACEHOLDERS = {"XXX", "https://github.com/SemiAnalysisAI/InferenceX/pull/XXX"}, and validate_added_pr_link (lines 142-145) raises ChangelogValidationError("new PR entry must use ... or an XXX placeholder; found 'TBD'") for any…

[chunfangamd]

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. https://github.com/SemiAnalysisAI/InferenceX/actions/runs/35873197630
• Verified that this PR passes evals. Please link to GitHub Action workflow that shows this. https://github.com/SemiAnalysisAI/InferenceX/actions/runs/35873197630
• 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 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
• 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 single-node vLLM/SGLang recipe in this PR is documented in the official vLLM recipes and/or the SGLang cookbook:
  • I linked the corresponding upstream PR in the vLLM recipe repo or SGLang 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:

• Upstream recipe (framework: atom, so the vLLM-recipes / SGLang-cookbook requirement is N/A): https://github.com/ROCm/ATOM/blob/main/recipes/MiniMax-M3-Agentic-InferenceX.md, which gained ATOM_PA_FLYDSL=1 in ROCm/ATOM#2366 (MERGED 2026-09-23). Major args match the TP4 arm: amd/MiniMax-M3-MXFP4, --kv_cache_dtype fp8, --block-size 128, --max-num-batched-tokens 32768, --max-num-seqs 2*CONC, the identical ptpc_fp8 --online_quant_config, thinking_mode: enabled, EAGLE3 Inferact/MiniMax-M3-EAGLE3-GQA with 3 draft tokens (acceptance rate 0.5933 = AL 2.78), ATOM_FORCE_ATTN_TRITON=1, AITER_QUICK_REDUCE_QUANTIZATION=INT4, ATOM_PA_FLYDSL=1, and the LMCache / ATOM_SLRU / all-rank lookup tier at C40+. InferenceX-only settings: the TP2 arms, cudagraph sizes, ATOM_M3_INDEXER_CP bands, NUMA-spread ROCR_VISIBLE_DEVICES, and AITER_SITUV2_A4W4 / AITER_FLYDSL_STAGE2_FP8 (target MoE kernels).

Draft-model precision record: MiniMax-M3 EAGLE3-GQA (Check 13)

• Draft and revision. The standalone PoR draft Inferact/MiniMax-M3-EAGLE3-GQA, HF revision 96692486b5fd38ebf8fd2a5f6bb53427d30819a8, loaded with --method eagle3 --draft-model Inferact/MiniMax-M3-EAGLE3-GQA --num-speculative-tokens 3. The draft is the same one used before this PR.
• Shipped precision: BF16. config.json is LlamaForCausalLMEagle3 with 1 decoder layer, hidden size 6144, torch_dtype: bfloat16, and no quantization_config. Its single model.safetensors holds 17 tensors (3.07B parameters), all BF16.
• Pinned-image default handling. The image is rocm/atom-dev:nightly_202609231248 (sha256:80f2c66579704aba933a9ff2a6e22e1dc61917c0fe6a2b0a0166816cd6d0100a), ATOM 0.1.7.dev16+g10ab5e8c1 (10ab5e8c1995c3e8d9e0674534f093d1e33a04ff, also printed in the eval job log). atom/models/eagle3_llama.py builds every draft linear (midlayer.self_attn.qkv_proj / o_proj, midlayer.mlp.gate_up_proj / down_proj, fc), plus lm_head and the embedding, without a quant_config. LinearBase then resolves LayerQuantConfig(), i.e. QuantType.No / BF16, and process_weights_after_loading online-quantizes only when self.quant_config is not None and self.quant_config.online_quant (atom/model_ops/linear.py). The ptpc_fp8 --online_quant_config therefore reaches only the target, and its exclude_layer list never applies to the draft.
• Effective precision: BF16 weights, same as shipped. The run 35873197630 server logs (TP2 C1 benchmark and TP4 C48 LMCache eval) show three steps. The target logs Post-load online quantization ... 126 layers online-quantized. The draft then logs weight load phases: ... (17 tensors) followed by Weight post-processing done: 0.00 seconds, with no online-quant pass. Finally it logs Eagle3 draft model loaded from Inferact/MiniMax-M3-EAGLE3-GQA (independent embed/lm_head), and the printed draft config reports dtype: bfloat16. The recipe passes no draft quantization or dtype flag.
• Draft KV cache: FP8, inherited from the target's --kv_cache_dtype fp8. The draft attention takes atom_config.kv_cache_dtype, so target and draft use the same upstream-supported KV dtype. The log line Allocated draft KV pool: 26949 blocks, 1821321216 B works out to 528 B/token at TP2, which is 2 KV heads × 128 × (K+V) × 1 B plus scales.
• ATOM_PA_FLYDSL=1 / ATOM_PA_FLYDSL_PLAN=1 (new in this PR) do not change precision. run_pa_decode (atom/model_ops/base_attention.py, ROCm/ATOM#2366) passes the same compute_type, q, k_scale / v_scale and FP8 cache to either kernel. It routes to FlyDSL only when the call is already compute_type == fp8 with an FP8 cache, and otherwise keeps gluon. compute_type comes from kv_cache_dtype in attention_mha.py, which #2366 left unchanged. The planner only changes how the KV is partitioned. The logs show pa_decode -> flydsl[...] (... head_dim=128 torch.float8_e4m3fn).
• Other recipe knobs. AITER_SITUV2_A4W4=1 and AITER_FLYDSL_STAGE2_FP8=1 are read only by aiter's MoE kernels (fused_moe.py, aot/flydsl/mxfp4_moe.py, ops/flydsl/moe_kernels.py), so they affect the target's MXFP4 MoE and not the dense draft. --index-cache-dtype fp8 applies to the target's sparse-attention indexer. AITER_QUICK_REDUCE_QUANTIZATION=INT4 has been in the recipe since #2816 and is also in the ATOM recipe; ATOM would set it automatically only for a llama target. It turns on aiter's INT4 quick all-reduce for the whole process, so TP all-reduces that exceed the fused custom-all-reduce limit are INT4-compressed in transit. That includes the draft's step-0 pass over a target prefill chunk. Decode-size draft all-reduces stay on the full-precision fused custom all-reduce. The setting does not change draft weights, GEMMs or attention.
• SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is not enabled. It is an SGLang variable that ATOM does not read. It is not set by the recipe script, benchmark_lib.sh, runtime_settings.sh or the MI355X launcher, nor in the launched environments of the benchmark and eval jobs.

Other checks

• Sweep and evals. Run 35873197630 ran on in-PR commit 7bbb0ec7; attempt 2 re-ran the two cells cancelled in attempt 1. All 19 agentic / cells succeeded on rocm/atom-dev:nightly_202609231248: TP4 C1-C32 and TP2 C1/C2 resident, plus TP2 C20/C25/C30 and TP4 C40/C48 with LMCache. So did the single agentic eval / cell (TP4 C48 LMCache, selected by eval-concs: [48]). minimax_m3_full em_strict is 0.97 (n=102, infrastructure_success: true) in eval_results_all/agg_eval_all.json. That is above the minimax_m3_full floor, since infx/evals/thresholds.yaml has no minimaxm3 override. The eval used real draft acceptance: its launch has no --spec-decode-acceptance-length. Reuse is set with /reuse-sweep-run 35873197630.
• PR head 2e4e7174. The head only merges main. The recipe script and the config block are byte-identical to 7bbb0ec7. The merged harness changes touch only the fixed-sequence and multi-node srt-slurm launch paths, GPU-telemetry CSV filtering, and a byte-identical move of wait_for_amd_gpu_clean. The head's Tests failures are the same 14 utils/test_srt_single_node.py::test_pool_launcher_stages_artifacts_and_propagates_failure[...] cases that fail on main 4f94801a; this PR's own tests pass.
• Golden AL. Throughput cells launch with --spec-decode-acceptance-length 2.78 --num-speculative-tokens 3 and thinking_mode: enabled; the server log shows Forced speculative acceptance ON: mean acceptance length 2.7800 over 3 draft positions. This matches golden_al_distribution/minimaxm3_eagle3_gqa.yaml minimax-m3.thinking_on[3] = 2.78.
• Pareto coverage. There is one affected curve: minimaxm3 / agentic-coding / MI355X / fp4 / ATOM / run 35873197630 / rocm/atom-dev:nightly_202609231248. It has 19 measured bmk_agentic_minimaxm3_* points, 0 invalid, and 8 on the tput_per_gpu vs p90_e2el frontier (TP4 C5, C15, C20, C24, C28, C32; TP2 LMCache C20, C25). The P75 frontier is 7/19. Reproduced with infx.workflows.pareto_coverage.
• Everything else. framework: atom, so the upstream vLLM/SGLang image rule does not apply; engine-first ordering holds through minimaxm3-fp4-mi355x-vllm-agentic-mtp on the same SKU. MiniMax-M3 agentic coding (EAGLE3) is active in MODELS.md, and Inferact/MiniMax-M3-EAGLE3-GQA is a PoR draft. There is no --hf-overrides. The target-only ptpc_fp8 online quant is target/verifier precision and is covered by the eval, and ATOM_M3_INDEXER_CP keeps selection bit-identical. AIPerf replays through --endpoint /v1/chat/completions --endpoint-type chat --apply-chat-template. There are no engine patches; the only install is a pip install --target of the sentencepiece / tiktoken tokenizer deps. The GPU-visibility change only removes the duplicate HIP mask. The new eval-concs changelog field (infx/matrix/plan.py, infx/matrix/validation.py) only filters evals that are already selected, and it rejects empty, invalid or conflicting selections. It is covered by new tests and EN/ZH docs. There is no append-only: true.

Signed: @chunfangamd

[github-actions]

✅✅✅ Verdict: PASS ✅✅✅

Passed and not applicable checks

✅ Check 0 (CODEOWNER): PASS — @chunfangamd is a listed owner of configs/amd-master.yaml; every other changed path falls only under the * catch-all.

✅ Check 1 (Passing sweep on in-PR commit): PASS — in-PR commit 7bbb0ec7 has 19/19 agentic / and 1/1 agentic eval / check-runs at success (run 35873197630, attempt 2). The pinned head 296a93ad only merges main: the recipe script and config block are byte-identical to 7bbb0ec7.

✅ Check 2 (Evals pass): PASS — agg_eval_all.json has minimax_m3_full em_strict 0.97, n=102, infrastructure_success: true (floor 0.0 in infx/evals/thresholds.yaml). The eval ran on rocm/atom-dev:nightly_202609231248, the PR image, with real acceptance.

➖ Check 3 (Recipe linked/merged): N/A — framework: atom, so the vLLM-recipes / SGLang-cookbook rule does not apply. For reference, the linked ROCm/ATOM#2366 is MERGED (2026-09-23).

✅ Check 4 (Reuse command): PASS — /reuse-sweep-run 35873197630 was posted by chunfangamd (COLLABORATOR).

✅ Check 5 (Latest checklist template): PASS — every item in the current docs/PR_REVIEW_CHECKLIST.md template is present and checked.

✅ Check 6 (Upstream images / engine-first): PASS — this is an ATOM entry, so the upstream-image rule is N/A. Engine-first ordering is satisfied by the existing minimaxm3-fp4-mi355x-vllm-agentic-mtp entry (same model and cluster:mi355x-amds runner, vllm/vllm-openai-rocm).

✅ Check 7 (No deprecated models/scenarios): PASS — minimaxm3 agentic coding with EAGLE3 is active in MODELS.md.

✅ Check 8 (No architecture hacks): PASS — no --hf-overrides or model-config edits. ptpc_fp8 online quant applies only to the target, and evals cover it.

✅ Check 9 (Spec-decode uses chat template): PASS — the replay adds --apply-chat-template on the chat endpoint, with thinking_mode: enabled.

✅ Check 10 (No engine patches): PASS — no patches, heredoc rewrites or engine wheels. The only install is a pip install --target of the sentencepiece/tiktoken tokenizer deps.

✅ Check 11 (Agentic golden AL): PASS — throughput cells pass --spec-decode-acceptance-length 2.78 with 3 draft tokens. The server log confirms Forced speculative acceptance ON: mean acceptance length 2.7800, which matches golden_al_distribution/minimaxm3_eagle3_gqa.yaml thinking_on[3] = 2.78. The eval runs without it, using real acceptance.

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

✅ Check 13 (Draft runs as shipped): PASS — the draft Inferact/MiniMax-M3-EAGLE3-GQA@96692486 has torch_dtype: bfloat16 and no quantization_config. ATOM eagle3_llama.py@10ab5e8c builds the draft layers with no quant_config. Logs show 126 layers online-quantized only for the target and Weight post-processing done: 0.00 seconds for the draft. There is no draft dtype or quant flag. The FP8 draft KV is inherited from the target's --kv_cache_dtype fp8. SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE does not appear in the recipe or the logs, and ATOM does not read it.

✅ Check 14 (Pareto coverage): PASS — there is one curve: minimaxm3 / agentic-coding / MI355X ATOM fp4 / rocm/atom-dev:nightly_202609231248, from run 35873197630 bmk_agentic_minimaxm3_* artifacts. It has 19 measured points and 0 invalid. The P90 tput_per_gpu vs p90_e2el frontier has 8 points (P75: 7), reproduced with infx.workflows.pareto_coverage.

Assessed commit: 296a93ad6c8fa017e83fb5f404f09f2efc16e1c4.

[functionstackx]

InferenceX has switched away from unmaintainable bash scripts to YAML files that don't repeat the same stuff over and over again. Please merge the latest main into this PR: we have migrated single-node AgentX onto native srt-slurm (#3428), so AgentX configs are now declarative YAML recipes, not per-config 1000+ line bash slop scripts. Please also delete the old benchmarks/single_node/** scripts (see this recipe for the new format).

[yhl-amd]

@functionstackx Merged latest main and moved this change onto the native srt-slurm recipe. The legacy benchmarks/single_node/agentic/minimaxm3_fp4_mi355x_atom_mtp.sh is deleted, and the PR now only touches benchmarks/single_node/srt-slurm-recipes/minimaxm3/atom/mi355x-fp4-mtp/agentic.yaml: bump the engine image to rocm/atom-dev:nightly_202609231248 (contains ROCm/ATOM#2366) and set ATOM_PA_FLYDSL=1 / ATOM_PA_FLYDSL_PLAN=1, plus the matching amd-master.yaml image, changelog entry and docs. The bash-only GPU-mask fix is dropped since it no longer applies.

@seungrokj I merged your 25bce60a on top of the migrated layout. TP4 C32 is dropped (along with its now-unused override_tp4_c32), and evals follow the default policy. The TP2 LMCache [20, 25] arm could not be carried over: #3428/#3461 removed all LMCache bands for this config because srtctl reserves ATOM's kv-transfer-config for disaggregated workers. It can come back once srt-slurm supports single-node ATOM LMCache.

Fresh sweep on head e027a054: run 36233199454 passed, 13/13 throughput (TP4 C1–C28, TP2 C1–C2) and 13/13 minimax_m3_full evals (em_strict 0.97–0.99, n=102 each). The server env confirms ATOM_PA_FLYDSL=1 ATOM_PA_FLYDSL_PLAN=1 on nightly_202609231248.