Validate archived MI355X MiniMax M3 vLLM disaggregation with srt-slurm

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/mi355x-fp4/8k1k/colocated-tp4-smoke.yaml

@@ -0,0 +1,151 @@
+# Smoke-only colocated variant of the archived MiniMax-M3 recipe.
+# Preserve the original TP4/TP4 serving and benchmark settings. srt-slurm
+# owns allocation, native vLLM Router discovery, health, logs, and teardown.
+schema: 2
+name: mi355x-vllm-minimaxm3-colocated-tp4-smoke
+
+model:
+  # Same provisioned checkpoint selected by the retired AMD launcher.
+  path: /it-share/data/MiniMax-M3-MXFP4
+  container: vllm/vllm-openai-rocm:v0.28.0
+  precision: fp4
+
+identity:
+  model:
+    repo: amd/MiniMax-M3-MXFP4
+  container:
+    image: vllm/vllm-openai-rocm:v0.28.0
+  frameworks:
+    vllm-router: 5d3c3af9a41dec2d5ecff412024b7a743763472d
+
+slurm:
+  time_limit: "08:00:00"
+
+resources:
+  gpu_type: mi355x
+  gpus_per_node: 8
+frontend:
+  type: vllm-router
+  container_image: vllm/vllm-router:nightly-20260922-5d3c3af
+  enable_multiple_frontends: false
+  args:
+    policy: consistent_hash
+    prefill-policy: consistent_hash
+    decode-policy: consistent_hash
+    log-level: info
+
+environment: &runtime_environment
+  HF_HOME: /hf_hub_cache
+  HF_HUB_CACHE: /hf_hub_cache
+  HUGGINGFACE_HUB_CACHE: /hf_hub_cache
+  PYTHONDONTWRITEBYTECODE: "1"
+  PYTHONUNBUFFERED: "1"
+  GLOO_SOCKET_IFNAME: eno0
+  NCCL_SOCKET_IFNAME: eno0
+  NCCL_IB_HCA: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
+  MORI_RDMA_TC: "104"
+  MORI_IO_SQ_BACKOFF_TIMEOUT_US: "50000"
+  MORI_IO_QP_MAX_SEND_WR: "16384"
+  MORI_IO_QP_MAX_CQE: "32768"
+  MORI_IO_QP_MAX_SGE: "2"
+  MORI_IO_TC_DISABLE: "0"
+  VLLM_USE_V1: "1"
+  VLLM_SERVER_DEV_MODE: "0"
+  VLLM_DISABLE_REQUEST_ID_RANDOMIZATION: "1"
+  VLLM_ROCM_USE_AITER: "1"
+  VLLM_ROCM_USE_AITER_MOE: "1"
+  VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: "1"
+  VLLM_USE_BREAKABLE_CUDAGRAPH: "0"
+  VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+
+engine:
+  type: vllm
+  connector: moriio
+roles:
+  prefill:
+    nodes: 1
+    workers: 1
+    gpus: 4
+    env:
+      VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: INT4
+      VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: "2048"
+    args: &server_options
+      tensor-parallel-size: 4
+      served-model-name: MiniMax-M3-MXFP4
+      max-num-batched-tokens: 32768
+      max-num-seqs: 512
+      block-size: 128
+      language-model-only: true
+      attention-backend: TRITON_ATTN
+      moe-backend: aiter
+      no-enable-prefix-caching: true
+      gpu-memory-utilization: 0.90
+      tool-call-parser: minimax_m3
+      reasoning-parser: minimax_m3
+      enable-auto-tool-choice: true
+      trust-remote-code: true
+  decode:
+    nodes: colocate
+    workers: 1
+    gpus: 4
+
+    args: *server_options
+
+# Reserve the whole node; each TP4 role owns a disjoint four-GPU subset.
+sbatch_directives:
+  cpus-per-task: "128"
+  mem: "0"
+  exclusive: ""
+
+srun_options:
+  container-writable: ""
+  container-remap-root: ""
+  mem: "0"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 5
+
+benchmark:
+  type: custom
+  command: |
+    set -eo pipefail
+    source /infmax-workspace/benchmarks/benchmark_lib.sh --validation-only
+    check_env_vars SLURM_JOB_ID SRT_FRONTEND_HOST SRT_FRONTEND_PORT CONC_LIST RANDOM_RANGE_RATIO
+    result_root="/results/${SLURM_JOB_ID}"
+    mkdir -p "${result_root}/fixed-seq"
+    archive_runtime_logs() {
+      tar -C /logs -czf "${result_root}/runtime-logs.tar.gz" . 2>/dev/null || true
+    }
+    trap archive_runtime_logs EXIT
+    for concurrency in ${CONC_LIST}; do
+      num_prompts=$((concurrency * 10))
+      if (( num_prompts < 16 )); then num_prompts=16; fi
+      python3 /infmax-workspace/utils/bench_serving/benchmark_serving.py \
+        --backend openai \
+        --base-url "http://${SRT_FRONTEND_HOST}:${SRT_FRONTEND_PORT}" \
+        --endpoint /v1/completions \
+        --model MiniMax-M3-MXFP4 \
+        --tokenizer /model \
+        --dataset-name random \
+        --random-input-len 8192 \
+        --random-output-len 1024 \
+        --random-range-ratio "${RANDOM_RANGE_RATIO}" \
+        --num-warmups "$((concurrency * 2))" \
+        --num-prompts "${num_prompts}" \
+        --max-concurrency "${concurrency}" \
+        --request-rate inf \
+        --ignore-eos \
+        --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result \
+        --result-dir "${result_root}/fixed-seq" \
+        --result-filename "minimaxm3-vllm-mi355x-disagg-1p1d-tp4-isl8192-osl1024-c${concurrency}.json"
+      # The original vLLM benchmark allowed the idle KV reaper ten seconds.
+      sleep 10
+    done
+  env:
+    <<: *runtime_environment
+    MODEL_NAME: MiniMax-M3-MXFP4
+    TRANSFORMERS_VERBOSITY: error
+    TOKENIZERS_PARALLELISM: "false"

benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/mi355x-fp4/8k1k/disagg-1p1d-tp4.yaml

@@ -0,0 +1,152 @@
+# Compatibility validation of the archived MI355X MiniMax-M3 FP4 recipe.
+# Preserve the original TP4/TP4 serving and benchmark settings. srt-slurm
+# owns allocation, native vLLM Router discovery, health, logs, and teardown.
+schema: 2
+name: mi355x-vllm-minimaxm3-disagg-1p1d-tp4-8k1k
+
+model:
+  # Same provisioned checkpoint selected by the retired AMD launcher.
+  path: /it-share/data/MiniMax-M3-MXFP4
+  container: vllm/vllm-openai-rocm:v0.28.0
+  precision: fp4
+
+identity:
+  model:
+    repo: amd/MiniMax-M3-MXFP4
+  container:
+    image: vllm/vllm-openai-rocm:v0.28.0
+  frameworks:
+    vllm-router: 5d3c3af9a41dec2d5ecff412024b7a743763472d
+
+slurm:
+  time_limit: "08:00:00"
+
+resources:
+  gpu_type: mi355x
+  gpus_per_node: 8
+frontend:
+  type: vllm-router
+  container_image: vllm/vllm-router:nightly-20260922-5d3c3af
+  enable_multiple_frontends: false
+  args:
+    policy: consistent_hash
+    prefill-policy: consistent_hash
+    decode-policy: consistent_hash
+    log-level: info
+
+environment: &runtime_environment
+  HF_HOME: /hf_hub_cache
+  HF_HUB_CACHE: /hf_hub_cache
+  HUGGINGFACE_HUB_CACHE: /hf_hub_cache
+  PYTHONDONTWRITEBYTECODE: "1"
+  PYTHONUNBUFFERED: "1"
+  GLOO_SOCKET_IFNAME: eno0
+  NCCL_SOCKET_IFNAME: eno0
+  NCCL_IB_HCA: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
+  MORI_RDMA_TC: "104"
+  MORI_IO_SQ_BACKOFF_TIMEOUT_US: "50000"
+  MORI_IO_QP_MAX_SEND_WR: "16384"
+  MORI_IO_QP_MAX_CQE: "32768"
+  MORI_IO_QP_MAX_SGE: "2"
+  MORI_IO_TC_DISABLE: "0"
+  VLLM_USE_V1: "1"
+  VLLM_SERVER_DEV_MODE: "0"
+  VLLM_DISABLE_REQUEST_ID_RANDOMIZATION: "1"
+  VLLM_ROCM_USE_AITER: "1"
+  VLLM_ROCM_USE_AITER_MOE: "1"
+  VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: "1"
+  VLLM_USE_BREAKABLE_CUDAGRAPH: "0"
+  VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+
+engine:
+  type: vllm
+  connector: moriio
+roles:
+  prefill:
+    nodes: 1
+    workers: 1
+    gpus: 4
+    env:
+      VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: INT4
+      VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: "2048"
+    args: &server_options
+      tensor-parallel-size: 4
+      served-model-name: MiniMax-M3-MXFP4
+      max-num-batched-tokens: 32768
+      max-num-seqs: 512
+      block-size: 128
+      language-model-only: true
+      attention-backend: TRITON_ATTN
+      moe-backend: aiter
+      no-enable-prefix-caching: true
+      gpu-memory-utilization: 0.90
+      tool-call-parser: minimax_m3
+      reasoning-parser: minimax_m3
+      enable-auto-tool-choice: true
+      trust-remote-code: true
+  decode:
+    nodes: 1
+    workers: 1
+    gpus: 4
+
+    args: *server_options
+
+# Preserve the original exclusive two-node allocation, although each worker
+# uses only four GPUs. Do not pack the roles onto one node.
+sbatch_directives:
+  cpus-per-task: "128"
+  mem: "0"
+  exclusive: ""
+
+srun_options:
+  container-writable: ""
+  container-remap-root: ""
+  mem: "0"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 5
+
+benchmark:
+  type: custom
+  command: |
+    set -eo pipefail
+    source /infmax-workspace/benchmarks/benchmark_lib.sh --validation-only
+    check_env_vars SLURM_JOB_ID SRT_FRONTEND_HOST SRT_FRONTEND_PORT CONC_LIST RANDOM_RANGE_RATIO
+    result_root="/results/${SLURM_JOB_ID}"
+    mkdir -p "${result_root}/fixed-seq"
+    archive_runtime_logs() {
+      tar -C /logs -czf "${result_root}/runtime-logs.tar.gz" . 2>/dev/null || true
+    }
+    trap archive_runtime_logs EXIT
+    for concurrency in ${CONC_LIST}; do
+      num_prompts=$((concurrency * 10))
+      if (( num_prompts < 16 )); then num_prompts=16; fi
+      python3 /infmax-workspace/utils/bench_serving/benchmark_serving.py \
+        --backend openai \
+        --base-url "http://${SRT_FRONTEND_HOST}:${SRT_FRONTEND_PORT}" \
+        --endpoint /v1/completions \
+        --model MiniMax-M3-MXFP4 \
+        --tokenizer /model \
+        --dataset-name random \
+        --random-input-len 8192 \
+        --random-output-len 1024 \
+        --random-range-ratio "${RANDOM_RANGE_RATIO}" \
+        --num-warmups "$((concurrency * 2))" \
+        --num-prompts "${num_prompts}" \
+        --max-concurrency "${concurrency}" \
+        --request-rate inf \
+        --ignore-eos \
+        --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result \
+        --result-dir "${result_root}/fixed-seq" \
+        --result-filename "minimaxm3-vllm-mi355x-disagg-1p1d-tp4-isl8192-osl1024-c${concurrency}.json"
+      # The original vLLM benchmark allowed the idle KV reaper ten seconds.
+      sleep 10
+    done
+  env:
+    <<: *runtime_environment
+    MODEL_NAME: MiniMax-M3-MXFP4
+    TRANSFORMERS_VERBOSITY: error
+    TOKENIZERS_PARALLELISM: "false"

configs/deprecated/amd-master.yaml

@@ -38,6 +38,39 @@ dsr1-fp4-mi355x-sglang-mtp:
       search-space:
       - { tp: 8, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp }
 
+minimaxm3-fp4-mi355x-vllm-disagg:
+  image: vllm/vllm-openai-rocm:v0.28.0
+  model: amd/MiniMax-M3-MXFP4
+  model-prefix: minimaxm3
+  runner: cluster:mi355x-amds
+  precision: fp4
+  framework: vllm-disagg
+  router: { name: vllm-router, version: "5d3c3af9a41dec2d5ecff412024b7a743763472d" }
+  kv-p2p-transfer: moriio
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - spec-decoding: "none"
+        conc-list: [1, 2, 4, 8, 16, 32, 64, 128, 256]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/mi355x-fp4/8k1k/disagg-1p1d-tp4.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/mi355x-fp4/8k1k/disagg-1p1d-tp4.yaml"
+
 dsr1-fp4-mi355x-atom:
   image: rocm/atom:rocm7.2.3_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom20260511
   model: amd/DeepSeek-R1-0528-MXFP4-Preview

configs/smoke/amd-srt-colocated.yaml

@@ -0,0 +1,36 @@
+# Smoke-only colocated layout; the archived two-node recipe remains unchanged.
+minimaxm3-fp4-mi355x-vllm-colocated-smoke:
+  image: vllm/vllm-openai-rocm:v0.28.0
+  model: amd/MiniMax-M3-MXFP4
+  model-prefix: minimaxm3
+  runner: cluster:mi355x-amds
+  precision: fp4
+  framework: vllm-disagg
+  router:
+    name: vllm-router
+    version: 5d3c3af9a41dec2d5ecff412024b7a743763472d
+  kv-p2p-transfer: moriio
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - spec-decoding: none
+        conc-list:
+        - 4
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - CONFIG_FILE=recipes/vllm/minimax-m3/mi355x-fp4/8k1k/colocated-tp4-smoke.yaml
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - CONFIG_FILE=recipes/vllm/minimax-m3/mi355x-fp4/8k1k/colocated-tp4-smoke.yaml