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

benchmarks/single_node/srt-slurm-recipes/minimaxm3/atom/mi355x-fp4-mtp/agentic.yaml

@@ -7,7 +7,7 @@ base:
   name: minimaxm3-fp4-mi355x-atom-agentic
   model:
     path: hf:amd/MiniMax-M3-MXFP4
-    container: rocm/atom-dev:nightly_202609171455
+    container: rocm/atom-dev:nightly_202609231248
     precision: fp4
   resources:
     gpu_type: mi355x
@@ -47,6 +47,8 @@ base:
         AITER_QUICK_REDUCE_QUANTIZATION: INT4
         AITER_FLYDSL_STAGE2_FP8: '1'
         ATOM_FORCE_ATTN_TRITON: '1'
+        ATOM_PA_FLYDSL: '1'
+        ATOM_PA_FLYDSL_PLAN: '1'
   benchmark:
     type: custom
     command: bash /infmax-workspace/benchmarks/srt_agentic.sh
@@ -57,7 +59,7 @@ base:
       AIPERF_APPLY_CHAT_TEMPLATE: 'true'
 
 # One variant per point. Admission is 2x CONC. TP4 shards the indexer across
-# ranks at the measured points (15, 20, 24, 28, 32).
+# ranks at the measured points (15, 20, 24, 28).
 override_tp4_c1:
   roles:
     agg:
@@ -176,18 +178,6 @@ override_tp4_c28:
     env:
       CONC: '28'
 
-override_tp4_c32:
-  roles:
-    agg:
-      gpus: 4
-      args:
-        max-num-seqs: 64
-      env:
-        ATOM_M3_INDEXER_CP: '1'
-  benchmark:
-    env:
-      CONC: '32'
-
 override_tp2_c1:
   roles:
     agg:

configs/amd-master.yaml

@@ -688,7 +688,7 @@ kimik3-fp4-mi355x-atom-agentic-mtp:
       - { tp: 8, kv-offloading: none, conc-list: [1, 4], spec-decoding: mtp, srt-recipe: benchmarks/single_node/srt-slurm-recipes/kimik3/atom/mi355x-fp4-mtp/agentic.yaml }
 
 minimaxm3-fp4-mi355x-atom-agentic-mtp:
-  image: rocm/atom-dev:nightly_202609171455
+  image: rocm/atom-dev:nightly_202609231248
   model: amd/MiniMax-M3-MXFP4
   model-prefix: minimaxm3
   runner: cluster:mi355x-amds
@@ -701,7 +701,7 @@ minimaxm3-fp4-mi355x-atom-agentic-mtp:
     # = TOTAL_CPU_DRAM_GB / TP = node_DRAM * dram-utilization / 8 (TP-independent).
     - dram-utilization: 0.687
       search-space:
-      - { tp: 4, kv-offloading: none, conc-list: [1, 2, 4, 5, 8, 10, 12, 15, 20, 24, 28, 32], spec-decoding: mtp, srt-recipe: benchmarks/single_node/srt-slurm-recipes/minimaxm3/atom/mi355x-fp4-mtp/agentic.yaml }
+      - { tp: 4, kv-offloading: none, conc-list: [1, 2, 4, 5, 8, 10, 12, 15, 20, 24, 28], spec-decoding: mtp, srt-recipe: benchmarks/single_node/srt-slurm-recipes/minimaxm3/atom/mi355x-fp4-mtp/agentic.yaml }
       - { tp: 2, kv-offloading: none, conc-list: [1, 2], spec-decoding: mtp, srt-recipe: benchmarks/single_node/srt-slurm-recipes/minimaxm3/atom/mi355x-fp4-mtp/agentic.yaml }
 
 dsr1-fp4-mi355x-sglang-disagg:

docs/configuration-procedures.md

@@ -293,6 +293,21 @@ The pinned image is the official ATOM nightly
 The recipe does not patch AITER source at runtime; TP communication
 fusion, DSpark K6 and graph capture use the implementation shipped in the image.
 
+### MiniMax-M3 ATOM FlyDSL paged decode
+
+`minimaxm3-fp4-mi355x-atom-agentic-mtp` uses
+`rocm/atom-dev:nightly_202609231248` with `ATOM_PA_FLYDSL=1` and
+`ATOM_PA_FLYDSL_PLAN=1`, following [ROCm/ATOM#2366](https://github.com/ROCm/ATOM/pull/2366)
+and the [upstream recipe](https://github.com/ROCm/ATOM/blob/94cde4ba786f45b38c26ee8201444659e44f861f/recipes/MiniMax-M3-Agentic-InferenceX.md).
+FlyDSL handles supported paged-decode shapes; its work planner balances dense
+decode by actual context length. Unsupported shapes retain the Gluon fallback.
+Verify the selected route and capture-time work-plan creation in `server.log`.
+
+The change is limited to the image and the two FlyDSL variables in
+`benchmarks/single_node/srt-slurm-recipes/minimaxm3/atom/mi355x-fp4-mtp/agentic.yaml`;
+TP4 C32 is dropped; the TP4 C1-C28 and TP2 C1-C2 points, EAGLE3 K3, golden AL 2.78 and indexer CP
+are unchanged.
+
 ### DeepSeek-V4.1-Flash DSpark
 
 The GB200 DSpark recipe uses a minimum CUDA graph capture size of 64 tokens to cover concurrent AgentX subagents. This raises c1/c2/c4 from 8/16/32 to 64; c8 and above retain their existing sizes. The full trace, AL 3.51, and Engram UVA settings are preserved; low-concurrency tail latency improvements require CI confirmation.

docs/configuration-procedures_zh.md

@@ -237,6 +237,21 @@ schedule 和 ragged verification 保持关闭。
 配方不再在运行时修改 AITER 源码；TP 通信融合、DSpark K6 和 graph capture
 直接使用镜像内实现。
 
+### MiniMax-M3 ATOM FlyDSL paged decode
+
+`minimaxm3-fp4-mi355x-atom-agentic-mtp` 按照
+[ROCm/ATOM#2366](https://github.com/ROCm/ATOM/pull/2366) 和
+[上游配方](https://github.com/ROCm/ATOM/blob/94cde4ba786f45b38c26ee8201444659e44f861f/recipes/MiniMax-M3-Agentic-InferenceX.md)，
+使用 `rocm/atom-dev:nightly_202609231248`，启用 `ATOM_PA_FLYDSL=1` 和
+`ATOM_PA_FLYDSL_PLAN=1`。FlyDSL 处理支持的 paged-decode shape，work planner
+按实际上下文长度均衡 dense decode 工作量；不支持的 shape 仍回退至 Gluon。
+从 `server.log` 核对实际路由，以及 work plan 是否在图捕获时创建。
+
+改动仅限
+`benchmarks/single_node/srt-slurm-recipes/minimaxm3/atom/mi355x-fp4-mtp/agentic.yaml`
+中的镜像和两个 FlyDSL 变量，并移除 TP4 C32；TP4 C1-C28、TP2 C1-C2、EAGLE3 K3、golden AL 2.78
+和 indexer CP 保持不变。
+
 ### DeepSeek-V4.1-Flash DSpark
 
 GB200 的 DSpark 配方将 CUDA graph 最小捕获范围设为 64 tokens，以覆盖 AgentX 子代理并发。这会将 c1/c2/c4 的上限从 8/16/32 提升至 64；c8 及以上保持原有大小。完整轨迹、AL 3.51 和 Engram UVA 配置保持不变；需通过 CI 验证低并发尾延迟改善。

perf-changelog.yaml

@@ -8962,3 +8962,13 @@
     - "旧镜像切出于 2026-09-16，新镜像领先 441 个提交，包含当前曲线所不含的三项 Kimi-K3 正确性修复：vllm-project/vllm#51483 不再将无状态的首个 chunk 当作 decode，vllm-project/vllm#57098 修复 Kimi-K3 reasoning parser，vllm-project/vllm#57430 支持 routed expert 量化。新镜像还包含本臂在并发大于 4 时均会使用的 ROCm CPU KV 卸载改动，包括 vllm-project/vllm#57160（ROCm CPU KV 卸载改用私有 pinned 张量）与 vllm-project/vllm#50045（卸载背压检测）。"
     - "This image bump does not change the DSpark draft model data type. Only the image: line changes and kimik3_fp4_mi355x_mtp.sh is unchanged; the draft loads unmodified from the published Inferact/Kimi-K3-DSpark checkpoint via --speculative-config (model=Inferact/Kimi-K3-DSpark, method=dspark). The only dtype in that speculative-config is kv_cache_dtype=fp8, which sets the draft KV-cache storage precision, not the draft weights. No flag overrides or re-quantizes the draft-model weights, so the draft dtype is preserved from its checkpoint across this re-sweep."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3419
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Update the MiniMax-M3 ATOM AgentX srt-slurm recipe from rocm/atom-dev:nightly_202609171455 to nightly_202609231248 (digest sha256:80f2c66579704aba933a9ff2a6e22e1dc61917c0fe6a2b0a0166816cd6d0100a), which contains ROCm/ATOM#2366, and set ATOM_PA_FLYDSL=1 and ATOM_PA_FLYDSL_PLAN=1 for FlyDSL paged decode with context-length-aware dense-decode work planning."
+    - "Search space, EAGLE3 K3 with golden AL 2.78, indexer CP, graph sizes and the atomesh frontend are unchanged. Drop TP4 concurrency 32 to bound the sweep; the search space is TP4 concurrency 1-28 and TP2 concurrency 1-2 with GPU-resident KV."
+    - "将 MiniMax-M3 ATOM AgentX srt-slurm 配方镜像从 nightly_202609171455 更新为包含 ROCm/ATOM#2366 的 nightly_202609231248，并设置 ATOM_PA_FLYDSL=1 与 ATOM_PA_FLYDSL_PLAN=1，启用 FlyDSL paged decode 及按上下文长度规划的 dense decode 工作划分。EAGLE3 K3（golden AL 2.78）、indexer CP、graph 档位与 atomesh 前端保持不变；移除 TP4 C32，搜索空间为 TP4 C1-C28 与 TP2 C1-C2（GPU 常驻 KV）。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3388