Enable MiniMax-M3 H100 FP8 indexer cache

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml

@@ -1,11 +1,11 @@
 # MiniMax-M3 MXFP8 AgentX on H100 with vLLM EAGLE3 (the GQA draft head) and
-# optional Mooncake DRAM KV offload. 26 GiB of weights per GPU live in host memory.
+# optional Mooncake DRAM KV offload and an FP8 MSA indexer cache.
 base:
   schema: 2
   name: minimaxm3-fp8-h100-vllm-agentic
   model:
     path: hf:MiniMaxAI/MiniMax-M3-MXFP8
-    container: vllm/vllm-openai:nightly-ddd6fbca148a867aad1fcab7ec72f582b9977db4
+    container: vllm/vllm-openai:nightly-36768d1bfd39094681cdbc8cb37d4b31c0729c89
     precision: fp8
   resources:
     gpu_type: h100
@@ -20,9 +20,7 @@ base:
   engine:
     type: vllm
     connector: null
-  # The engine may take the full VLLM_ENGINE_READY_TIMEOUT_S to become ready.
-  # Lazy loads of the 31 shards from shared NFS ran at ~105 s/shard, past the
-  # script's 3600 s, so both windows are two hours.
+  # Allow up to two hours for model loading and engine readiness.
   health_check:
     interval_seconds: 10
     max_attempts: 720
@@ -35,11 +33,12 @@ base:
         served-model-name: MiniMaxAI/MiniMax-M3-MXFP8
         tensor-parallel-size: 8
         data-parallel-size: 1
-        gpu-memory-utilization: 0.90
-        cpu-offload-gb: 26
+        gpu-memory-utilization: 0.93
+        cpu-offload-gb: 5
         attention-backend: TRITON_ATTN
         safetensors-load-strategy: lazy
         kv-cache-dtype: fp8
+        attention-config: '{"indexer_kv_dtype":"fp8"}'
         block-size: 128
         language-model-only: true
         enable-prefix-caching: true
@@ -55,6 +54,7 @@ base:
       env:
         PYTHONNOUSERSITE: '1'
         VLLM_ENGINE_READY_TIMEOUT_S: '7200'
+        VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS: '0'
   benchmark:
     type: custom
     command: bash /infmax-workspace/benchmarks/srt_agentic.sh
@@ -121,8 +121,8 @@ override_tp8_c5:
       KV_OFFLOADING: none
 
 # DRAM points offload KV to an embedded Mooncake store. Per rank: the host budget
-# (1731 GB = 1612 GiB) less the 414 GiB checkpoint page cache, over TP8, less the
-# 26 GiB CPU weight offload and the 4 GiB local buffer = 119 GB.
+# (1676 GB = 1561 GiB) less the 414 GiB checkpoint page cache, over TP8, less the
+# 5 GiB CPU weight offload and the 4 GiB local buffer = 134 GB.
 override_tp8_c6_dram:
   # The worker's Mooncake client and the master run the same pinned release.
   setup_script: vllm-mooncake-0.3.11.sh
@@ -144,9 +144,9 @@ override_tp8_c6_dram:
         store_config:
           mode: embedded
           metadata_server: P2PHANDSHAKE
-          global_segment_size: 119GB
+          global_segment_size: 134GB
           local_buffer_size: 4GB
-          protocol: rdma
+          protocol: tcp
           device_name: ''
           enable_offload: false
   roles:
@@ -164,7 +164,7 @@ override_tp8_c6_dram:
     env:
       CONC: '6'
       KV_OFFLOADING: dram
-      TOTAL_CPU_DRAM_GB: '1731'
+      TOTAL_CPU_DRAM_GB: '1676'
 
 override_tp8_c8_dram:
   # The worker's Mooncake client and the master run the same pinned release.
@@ -187,9 +187,9 @@ override_tp8_c8_dram:
         store_config:
           mode: embedded
           metadata_server: P2PHANDSHAKE
-          global_segment_size: 119GB
+          global_segment_size: 134GB
           local_buffer_size: 4GB
-          protocol: rdma
+          protocol: tcp
           device_name: ''
           enable_offload: false
   roles:
@@ -207,4 +207,4 @@ override_tp8_c8_dram:
     env:
       CONC: '8'
       KV_OFFLOADING: dram
-      TOTAL_CPU_DRAM_GB: '1731'
+      TOTAL_CPU_DRAM_GB: '1676'

inferencex-e2e/configs/nvidia-master.yaml

@@ -5367,10 +5367,10 @@ qwen3.5-fp4-b200-trt-mtp:
         srt-recipe: benchmarks/single_node/srt-slurm-recipes/qwen3.5/trtllm/b200-fp4-mtp/8k1k.yaml
 
 minimaxm3-fp8-h100-vllm-agentic-mtp:
-  image: vllm/vllm-openai:nightly-ddd6fbca148a867aad1fcab7ec72f582b9977db4
+  image: vllm/vllm-openai:nightly-36768d1bfd39094681cdbc8cb37d4b31c0729c89
   model: MiniMaxAI/MiniMax-M3-MXFP8
   model-prefix: minimaxm3
-  runner: cluster:h100-dgxc
+  runner: cluster:h100-dsxe
   precision: fp8
   framework: vllm
   multinode: false

inferencex-e2e/configs/runners.yaml

@@ -2,6 +2,9 @@ labels:
   h100:
     - h100-cw_00
     - h100-cw_01
+    - h100-dgxc-new_00
+    - h100-dgxc-new_01
+    - h100-dgxc-new_02
     - h100-dgxc-slurm_00
     - h100-dgxc-slurm_01
     - h100-dgxc-slurm_02
@@ -156,6 +159,14 @@ labels:
   cluster:h100-cw:
     - h100-cw_00
     - h100-cw_01
+  h100-dgxc-new:
+    - h100-dgxc-new_00
+    - h100-dgxc-new_01
+    - h100-dgxc-new_02
+  cluster:h100-dsxe:
+    - h100-dgxc-new_00
+    - h100-dgxc-new_01
+    - h100-dgxc-new_02
   cluster:h100-dgxc:
     - h100-dgxc-slurm_00
     - h100-dgxc-slurm_01
@@ -326,6 +337,36 @@ clusters:
         network-interface: ""
         gpus-per-node-directive: false
         single-node-exclusive: false
+  h100-dsxe:
+    gpus-per-node: 8
+    available-cpu-dram-mib: 1_998_848
+    arch: x86_64
+    models:
+      entries:
+        MiniMax-M3-MXFP8: {root: data-models, dir: MiniMax-M3-MXFP8}
+    scheduler: slurm
+    slurm:
+      partition: batch_1
+      account: benchmark
+      exclusive: false
+      gres: "gpu:{gpus}"
+      volumes:
+        hf-hub-cache: {path: /data/home/sa-gha-runner/gharunners/hf-hub-cache}
+        aiperf-cache: {path: /data/home/sa-gha-runner/gharunners/ai-perf-cache}
+        data-models: {path: /data/models}
+      squash:
+        dir: /data/home/sa-gha-runner/gharunners/squash
+        visibility: shared
+        import: compute
+        lock-timeout-s: 600
+      srt-slurm:
+        extra:
+          default_gpu_exporter:
+            container_image: nvcr.io#nvidia/k8s/dcgm-exporter:4.6.0-4.8.3-distroless
+            port: 9401
+            command: "dcgm-exporter --collect-interval=1000 --address :{port} -f /configs/dcgm-counters-noprof.csv"
+        network-interface: ""
+        single-node-models: staged
   h100-dgxc:
     gpus-per-node: 8
     available-cpu-dram-mib: 2_063_837

inferencex-e2e/perf-changelog.yaml

@@ -9180,3 +9180,39 @@
     - "Update the SGLang ROCm image from lmsysorg/sglang-rocm:v0.5.20-rocm720-mi35x-20260927 to lmsysorg/sglang-rocm:v0.5.20-rocm720-mi35x-20260929 on this arm only. The other MI355X arms are left on their current tags."
     - "No serving flag outside the HiCache block changes, and no other config key is touched. Each of the 19 matrix points resolves to exactly one recipe override, with no override left unused."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3611
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Update the MiniMax-M3 H100 vLLM image, enable the FP8 MSA indexer cache, adjust CPU weight offload and GPU memory settings, and resize the Mooncake store segments."
+    - "更新 MiniMax-M3 H100 的 vLLM 镜像，启用 FP8 MSA 索引缓存，调整 CPU 权重卸载与 GPU 内存设置，并调整 Mooncake 存储段大小。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Route the MiniMax-M3 H100 AgentX config to cluster:h100-cw and align the Mooncake DRAM store size with that cluster's host-memory budget."
+    - "将 MiniMax-M3 H100 AgentX 配置切换到 cluster:h100-cw，并根据该集群的主机内存预算调整 Mooncake DRAM 存储段大小。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Route the MiniMax-M3 H100 AgentX config to the DSXE runner group and add its Slurm cluster profile."
+    - "将 MiniMax-M3 H100 AgentX 配置切换到 DSXE 运行器组，并添加对应的 Slurm 集群配置。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Use TCP transport for the single-node Mooncake DRAM store in the concurrency 6 and 8 variants."
+    - "将并发数为 6 和 8 的单节点 Mooncake DRAM 存储变体切换为 TCP 传输。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620