MiniMax-M3 MXFP8 full sweep config for GB300

.github/configs/nvidia-master.yaml

@@ -11679,6 +11679,128 @@ minimaxm2.5-fp8-gb300-dynamo-vllm:
           ep: 4
           dp-attn: true
 
+# MiniMax-M3 GB300 full sweep — safetensors-load-strategy removed from all
+# GB300 recipes (host-memory OOM with prefetch on CW Grace Blackwell nodes).
+# srun_options mem=0 required (DefMemPerCPU=4096 cgroup limit).
+minimaxm3-fp8-gb300-dynamo-vllm:
+  image: vllm/vllm-openai:minimax-m3
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: gb300-cw
+  precision: fp8
+  framework: dynamo-vllm
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      # Low latency: TP=4 aggregated, 1 node (4 GPU).
+      - conc-list: [4, 8, 16, 32, 64]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/agg-gb300-tp4-1n.yaml"
+        decode:
+          num-worker: 0
+          tp: 4
+          ep: 1
+          dp-attn: false
+
+      # Low latency: TP=8 aggregated, 2 nodes (8 GPU).
+      - conc-list: [4, 8, 16, 32, 64]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/agg-gb300-tp8-2n.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          ep: 1
+          dp-attn: false
+
+      # Mid curve: TP4+EP4 aggregated, 1 node (4 GPU).
+      - conc-list: [128, 256, 512]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/agg-gb300-tp4ep4-1n.yaml"
+        decode:
+          num-worker: 0
+          tp: 4
+          ep: 4
+          dp-attn: false
+
+      # Mid curve: 1P+1D disagg TP4+EP4, 2 nodes (4 GPU each).
+      - conc-list: [64, 128, 256, 512]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-tp4ep4-2n.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: false
+
+      # Mid curve: 1P+1D disagg TP4+EP4, collocated 1 node (8 GPU).
+      - conc-list: [64, 128, 256, 512]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-tp4ep4-1n.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: false
+
+      # High throughput: DEP=4 aggregated, 1 node (4 GPU).
+      - conc-list: [256, 512, 1024]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/agg-gb300-dep4-1n.yaml"
+        decode:
+          num-worker: 0
+          tp: 1
+          ep: 4
+          dp-attn: true
+
+      # Max throughput: DEP=8 aggregated, 2 nodes (8 GPU).
+      - conc-list: [512, 1024, 2048]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/agg-gb300-dep8-2n.yaml"
+        decode:
+          num-worker: 0
+          tp: 1
+          ep: 8
+          dp-attn: true
+
 # MiniMax-M3 day-zero (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3).
 # 427B total / 26B active MoE with MSA sparse attention; MXFP8 checkpoint
 # (MiniMaxAI/MiniMax-M3-MXFP8, ~444 GB) quantized by NVIDIA — native MX tensor

.github/workflows/benchmark-multinode-tmpl.yml

@@ -123,6 +123,11 @@ on:
 
 env:
   RANDOM_RANGE_RATIO: 0.8
+  # Day-zero models resolved via hf: ids download from the Hub inside the
+  # slurm job (srtctl pre-download + dynamo hub fetch). Anonymous requests
+  # get 429-rate-limited when several workers pull a 444 GB snapshot at
+  # once; sbatch/srun inherit this env so the token reaches the workers.
+  HF_TOKEN: ${{ secrets.INFERENCEX_OFFICIAL_RO_HF_TOKEN }}
   EXP_NAME: ${{ inputs.exp-name }}
   IMAGE: ${{ inputs.image }}
   MODEL_PREFIX: ${{ inputs.model-prefix }}

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/agg-gb300-dep4-1n.yaml

@@ -0,0 +1,91 @@
+name: "minimax-m3-vllm-agg-gb300-dep4-1n-1k1k"
+
+# MiniMax-M3 day-zero aggregated recipe
+# (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3). Aggregated-only on
+# day zero: NixlConnector KV transfer has never been exercised against
+# M3's MSA index cache, so disagg shapes are deferred until agg
+# baselines exist. model.path uses the hf: prefix — srtctl pre-downloads
+# the 444 GB MXFP8 snapshot once, on a compute node, into the shared
+# HF_HOME below (launcher seds __M3_HF_HOME__ to a cluster-writable
+# shared-FS path and bind-mounts it via extra_mount).
+# --block-size 128 is mandatory (MSA sparse/index cache alignment); the
+# benchmark is text-only, so language-model-only skips the vision
+# encoder. max-model-len / cudagraph capture / batched tokens are
+# trimmed to the fixed-seq-len scenario instead of the 1M default.
+
+model:
+  path: "hf:MiniMaxAI/MiniMax-M3-MXFP8"
+  container: "vllm/vllm-openai:minimax-m3"
+  precision: "fp8"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+# CW gb300 has no per-GPU Slurm defaults: without sbatch-level mem=0 the
+# job allocation is ntasks x DefMemPerCPU = 4 GB/node and worker cgroups
+# OOM-kill during engine init; srun_options.mem=0 alone only grants a
+# step what the job already holds. cpus-per-task=72 (one NUMA socket)
+# keeps the gpu-less infra step (etcd/nats) off the 1-CPU default.
+# Full rationale: vllm/deepseek-v4/agentic/disagg-gb300-1p6d-dep4-tp4-agentic.yaml.
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+srun_options:
+  mem: "0"
+
+
+slurm:
+  time_limit: "8:00:00"
+
+# 444 GB of weights off shared FS (cold HF cache on the first run):
+# allow up to 2 h for engine readiness.
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+extra_mount:
+  - "__M3_HF_HOME__:__M3_HF_HOME__"
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  agg_nodes: 1
+  agg_workers: 1
+  gpus_per_agg: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  aggregated_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    HF_HOME: "__M3_HF_HOME__"
+
+  vllm_config:
+    aggregated:
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      max-model-len: 2304
+      max-num-batched-tokens: 2048
+      max-cudagraph-capture-size: 512
+      block-size: 128
+      language-model-only: true
+      gpu-memory-utilization: 0.9
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      stream-interval: 128
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "256x512x1024"

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/agg-gb300-dep8-2n.yaml

@@ -0,0 +1,91 @@
+name: "minimax-m3-vllm-agg-gb300-dep8-2n-1k1k"
+
+# MiniMax-M3 day-zero aggregated recipe
+# (https://recipes.vllm.ai/MiniMaxAI/MiniMax-M3). Aggregated-only on
+# day zero: NixlConnector KV transfer has never been exercised against
+# M3's MSA index cache, so disagg shapes are deferred until agg
+# baselines exist. model.path uses the hf: prefix — srtctl pre-downloads
+# the 444 GB MXFP8 snapshot once, on a compute node, into the shared
+# HF_HOME below (launcher seds __M3_HF_HOME__ to a cluster-writable
+# shared-FS path and bind-mounts it via extra_mount).
+# --block-size 128 is mandatory (MSA sparse/index cache alignment); the
+# benchmark is text-only, so language-model-only skips the vision
+# encoder. max-model-len / cudagraph capture / batched tokens are
+# trimmed to the fixed-seq-len scenario instead of the 1M default.
+
+model:
+  path: "hf:MiniMaxAI/MiniMax-M3-MXFP8"
+  container: "vllm/vllm-openai:minimax-m3"
+  precision: "fp8"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+# CW gb300 has no per-GPU Slurm defaults: without sbatch-level mem=0 the
+# job allocation is ntasks x DefMemPerCPU = 4 GB/node and worker cgroups
+# OOM-kill during engine init; srun_options.mem=0 alone only grants a
+# step what the job already holds. cpus-per-task=72 (one NUMA socket)
+# keeps the gpu-less infra step (etcd/nats) off the 1-CPU default.
+# Full rationale: vllm/deepseek-v4/agentic/disagg-gb300-1p6d-dep4-tp4-agentic.yaml.
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+srun_options:
+  mem: "0"
+
+
+slurm:
+  time_limit: "8:00:00"
+
+# 444 GB of weights off shared FS (cold HF cache on the first run):
+# allow up to 2 h for engine readiness.
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+extra_mount:
+  - "__M3_HF_HOME__:__M3_HF_HOME__"
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 8
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  aggregated_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    HF_HOME: "__M3_HF_HOME__"
+
+  vllm_config:
+    aggregated:
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      max-model-len: 2304
+      max-num-batched-tokens: 2048
+      max-cudagraph-capture-size: 512
+      block-size: 128
+      language-model-only: true
+      gpu-memory-utilization: 0.9
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      stream-interval: 128
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "512x1024x2048"