[AgentX] MiniMax-M3.1-Flash B300 SGLang: add TP2 configs / [AgentX] MiniMax-M3.1-Flash B300 SGLang：新增 TP2 配置

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm31flash/sglang/b300-fp4-mtp/agentic-tp2.yaml

@@ -0,0 +1,89 @@
+base:
+  schema: 2
+  name: minimaxm31flash-fp4-b300-sglang-agentic-tp2
+  resources:
+    gpu_type: b300
+    gpus_per_node: 8
+  frontend:
+    type: sglang
+    enable_multiple_frontends: false
+  observability:
+    enabled: false
+    tachometer:
+      enabled: false
+  engine: sglang
+  # Cold loads of the 250 GB checkpoint plus JIT MSA kernel builds and graph capture outlast the default.
+  health_check:
+    max_attempts: 1440
+    interval_seconds: 5
+  roles:
+    agg:
+      nodes: 1
+      workers: 1
+      gpus: 2
+      # Serve flags follow the cookbook's B300 cells at
+      # https://docs.sglang.io/cookbook/autoregressive/MiniMax/MiniMax-M3.1.
+      args:
+        served-model-name: MiniMaxAI/MiniMax-M3.1-Flash
+        trust-remote-code: true
+        reasoning-parser: minimax-m3
+        tool-call-parser: minimax-m3
+        # Seven-token DSpark, the bundled draft's dspark_block_size. Throughput runs
+        # switch to synthetic acceptance at the golden acceptance length.
+        speculative-algorithm: DSPARK
+        speculative-dspark-block-size: 7
+        speculative-use-block-verification: true
+        # FP8 draft KV, as vLLM's DSpark drafter uses; bf16 costs ~10 GB of HBM per rank.
+        speculative-draft-kv-cache-dtype: fp8_e4m3
+        cuda-graph-backend-prefill: breakable
+        tokenizer-worker-num: 8
+        watchdog-timeout: 3600
+        enable-metrics: true
+      env:
+        PYTHONNOUSERSITE: '1'
+        PYTHONUNBUFFERED: '1'
+        SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+        FMHA_SM100_ALLOW_JIT: '1'
+  benchmark:
+    env:
+      # Warmup can leave bytes unacknowledged past AIPerf's 30 s default.
+      AIPERF_HTTP_TCP_USER_TIMEOUT: '900000'
+      AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:'
+
+zip_override_tp2_hicache:
+  roles:
+    agg:
+      args:
+        tensor-parallel-size: 2
+        moe-runner-backend: flashinfer_trtllm_routed
+        chunked-prefill-size: 32768
+        # 2x CONC (AgentX subagent fan-out); the decode graph batch matches it.
+        max-running-requests: [64, 72, 80]
+        cuda-graph-max-bs-decode: [64, 72, 80]
+        mem-fraction-static: 0.87
+        enable-hierarchical-cache: true
+        hicache-size: 370
+        hicache-write-policy: write_back
+  benchmark:
+    env:
+      CONC: ['32', '36', '40']
+      KV_OFFLOADING: dram
+
+zip_override_tp2_hicache_lpm:
+  roles:
+    agg:
+      args:
+        tensor-parallel-size: 2
+        moe-runner-backend: flashinfer_trtllm_routed
+        chunked-prefill-size: 32768
+        max-running-requests: [88, 92, 96]
+        cuda-graph-max-bs-decode: [88, 92, 96]
+        mem-fraction-static: 0.87
+        enable-hierarchical-cache: true
+        hicache-size: 370
+        hicache-write-policy: write_back
+        schedule-policy: lpm
+  benchmark:
+    env:
+      CONC: ['44', '46', '48']
+      KV_OFFLOADING: dram

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm31flash/sglang/b300-fp4-mtp/agentic.yaml

@@ -33,6 +33,8 @@ base:
         speculative-algorithm: DSPARK
         speculative-dspark-block-size: 7
         speculative-use-block-verification: true
+        # FP8 draft KV, as vLLM's DSpark drafter uses; bf16 costs ~10 GB of HBM per rank.
+        speculative-draft-kv-cache-dtype: fp8_e4m3
         cuda-graph-backend-prefill: breakable
         tokenizer-worker-num: 8
         watchdog-timeout: 3600

inferencex-e2e/configs/nvidia-master.yaml

@@ -9713,6 +9713,9 @@ minimaxm31flash-fp4-b300-sglang-agentic-dspark:
     - dram-utilization: 0.80
       search-space:
       - { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [32, 64, 80, 96], srt-recipe: agentic.yaml }
+    - dram-utilization: 1.0
+      search-space:
+      - { tp: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [32, 36, 40, 44, 46, 48], srt-recipe: agentic-tp2.yaml }
 
 # Official SGLang nightly; TP4 C1 uses GPU Engram, other points use host DRAM.
 dsv41flash-fp4-gb300-sglang-agentic-dspark:

inferencex-e2e/perf-changelog.yaml

@@ -11556,3 +11556,12 @@
     - "Disagg: 6 DEP2 prefill workers with a DEP8 decode at conc 280-360 or a DEP16 decode at conc 300; DeepGEMM MegaMoE, NIXL transfer plus a shared MooncakeStore DRAM pool (160 GB per rank), Dynamo KV router."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4468
   scenario-type: [agentic-coding]
+
+- config-keys:
+    - minimaxm31flash-fp4-b300-sglang-agentic-dspark
+  description:
+    - "Add TP2 points with DRAM KV offload (L2 HiCache) at conc 32 to 48."
+    - "Store the DSpark draft KV cache in FP8 (speculative-draft-kv-cache-dtype fp8_e4m3) on every point instead of the bf16 default, matching the vLLM DSpark drafter; this frees about 10 GB of HBM per rank and about 35 GB of the HiCache host pool for target KV."
+    - "Rerun the whole curve so every point uses the FP8 draft KV cache."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4524
+  scenario-type: [agentic-coding]