feat: MiniMax M2.5 PD Disaggregation Recipe (1P2D, MoRI-EP + MoRI-IO)

.github/configs/amd-master.yaml

@@ -1031,6 +1031,156 @@ dsr1-fp8-mi355x-sglang-disagg-mtp:
   #       - "DECODE_MTP_SIZE=0"
 
 
+dsr1-fp8-mi355x-vllm-disagg:
+  image: vllm/vllm-openai-rocm:v0.17.1
+  model: deepseek-ai/DeepSeek-R1-0528
+  model-prefix: dsr1
+  runner: mi355x-disagg
+  precision: fp8
+  framework: vllm-disagg
+  multinode: true
+  disagg: true
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    # 1P2D: 1 prefill node (co-located with proxy) + 2 decode nodes = 3 nodes total
+    - spec-decoding: "none"
+      conc-list: [ 8, 16, 32, 64, 128, 256, 512 ]
+      prefill:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: false
+        additional-settings:
+        - "PREFILL_NODES=1"
+        - "VLLM_MORIIO_CONNECTOR_READ_MODE=1"
+      decode:
+        num-worker: 2
+        tp: 8
+        ep: 8
+        dp-attn: false
+        additional-settings:
+        - "DECODE_NODES=2"
+
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - spec-decoding: "none"
+      conc-list: [ 8, 16, 32, 64, 128, 256, 512 ]
+      prefill:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: false
+        additional-settings:
+        - "PREFILL_NODES=1"
+        - "VLLM_MORIIO_CONNECTOR_READ_MODE=1"
+      decode:
+        num-worker: 2
+        tp: 8
+        ep: 8
+        dp-attn: false
+        additional-settings:
+        - "DECODE_NODES=2"
+
+  - isl: 1024
+    osl: 8192
+    search-space:
+    - spec-decoding: "none"
+      conc-list: [ 8, 16, 32, 64, 128, 256, 512 ]
+      prefill:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: false
+        additional-settings:
+        - "PREFILL_NODES=1"
+        - "VLLM_MORIIO_CONNECTOR_READ_MODE=1"
+      decode:
+        num-worker: 2
+        tp: 8
+        ep: 8
+        dp-attn: false
+        additional-settings:
+        - "DECODE_NODES=2"
+
+
+minimaxm25-fp8-mi355x-vllm-disagg:
+  image: minimax-m25-disagg:latest  # custom build: docker/minimax-m25-disagg/build.sh
+  model: MiniMaxAI/MiniMax-M2.5
+  model-prefix: minimaxm25
+  runner: mi355x-disagg
+  precision: fp8
+  framework: vllm-disagg
+  multinode: true
+  disagg: true
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    # 1P2D: 1 prefill node (co-located with proxy) + 2 decode nodes = 3 nodes total
+    - spec-decoding: "none"
+      conc-list: [ 8, 16, 32, 64, 128, 256, 512 ]
+      prefill:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: false
+        additional-settings:
+        - "PREFILL_NODES=1"
+        - "VLLM_MORIIO_CONNECTOR_READ_MODE=1"
+      decode:
+        num-worker: 2
+        tp: 8
+        ep: 8
+        dp-attn: false
+        additional-settings:
+        - "DECODE_NODES=2"
+
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - spec-decoding: "none"
+      conc-list: [ 8, 16, 32, 64, 128, 256, 512 ]
+      prefill:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: false
+        additional-settings:
+        - "PREFILL_NODES=1"
+        - "VLLM_MORIIO_CONNECTOR_READ_MODE=1"
+      decode:
+        num-worker: 2
+        tp: 8
+        ep: 8
+        dp-attn: false
+        additional-settings:
+        - "DECODE_NODES=2"
+
+  - isl: 1024
+    osl: 8192
+    search-space:
+    - spec-decoding: "none"
+      conc-list: [ 8, 16, 32, 64, 128, 256, 512 ]
+      prefill:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: false
+        additional-settings:
+        - "PREFILL_NODES=1"
+        - "VLLM_MORIIO_CONNECTOR_READ_MODE=1"
+      decode:
+        num-worker: 2
+        tp: 8
+        ep: 8
+        dp-attn: false
+        additional-settings:
+        - "DECODE_NODES=2"
+
+
 dsr1-fp4-mi355x-sglang-disagg:
   image: rocm/sgl-dev:sglang-0.5.9-rocm720-mi35x-mori-0227-3
   model: amd/DeepSeek-R1-0528-MXFP4

benchmarks/multi_node/dsr1_fp8_mi355x_vllm-disagg.sh

@@ -0,0 +1,79 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    CONC_LIST \
+    ISL \
+    OSL \
+    IMAGE \
+    SPEC_DECODING \
+    MODEL_PATH \
+    PREFILL_NUM_WORKERS \
+    PREFILL_TP \
+    PREFILL_EP \
+    PREFILL_DP_ATTN \
+    DECODE_NUM_WORKERS \
+    DECODE_TP \
+    DECODE_EP \
+    DECODE_DP_ATTN \
+    PREFILL_NODES \
+    DECODE_NODES \
+    RANDOM_RANGE_RATIO
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+set -x
+
+cd "$GITHUB_WORKSPACE/benchmarks/multi_node/vllm_disagg_utils" || exit 1
+
+export TIME_LIMIT="08:00:00"
+export MODEL_PATH=$MODEL_PATH
+export MODEL_NAME=$MODEL_NAME
+export CONTAINER_IMAGE=$IMAGE
+
+# Same EP/DP booleans as dsr1_fp8_mi355x_sglang-disagg.sh → amd_utils/submit.sh
+if [[ "${PREFILL_EP:-1}" -eq 1 ]]; then
+    export PREFILL_ENABLE_EP=false
+else
+    export PREFILL_ENABLE_EP=true
+fi
+
+if [[ "$PREFILL_DP_ATTN" == "true" ]]; then
+    export PREFILL_ENABLE_DP=true
+else
+    export PREFILL_ENABLE_DP=false
+fi
+
+if [[ "${DECODE_EP:-1}" -eq 1 ]]; then
+    export DECODE_ENABLE_EP=false
+else
+    export DECODE_ENABLE_EP=true
+fi
+
+if [[ "$DECODE_DP_ATTN" == "true" ]]; then
+    export DECODE_ENABLE_DP=true
+else
+    export DECODE_ENABLE_DP=false
+fi
+
+# Parameter order matches SGLang disagg submit.sh; arg 16 is optional NODELIST.
+JOB_ID=$(bash ./submit.sh $PREFILL_NODES \
+    $PREFILL_NUM_WORKERS \
+    $DECODE_NODES \
+    $DECODE_NUM_WORKERS \
+    $ISL $OSL "${CONC_LIST// /x}" inf \
+    ${PREFILL_ENABLE_EP} ${PREFILL_ENABLE_DP} \
+    ${DECODE_ENABLE_EP} ${DECODE_ENABLE_DP} \
+    ${PREFILL_TP} ${DECODE_TP} \
+    ${RANDOM_RANGE_RATIO} \
+    "${NODELIST:-}")
+
+if [[ $? -ne 0 ]]; then
+    echo "Failed to submit job" >&2
+    exit 1
+fi
+
+echo "$JOB_ID"

benchmarks/multi_node/minimaxm25_fp8_mi355x_vllm-disagg.sh

@@ -0,0 +1,77 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    CONC_LIST \
+    ISL \
+    OSL \
+    IMAGE \
+    SPEC_DECODING \
+    MODEL_PATH \
+    PREFILL_NUM_WORKERS \
+    PREFILL_TP \
+    PREFILL_EP \
+    PREFILL_DP_ATTN \
+    DECODE_NUM_WORKERS \
+    DECODE_TP \
+    DECODE_EP \
+    DECODE_DP_ATTN \
+    PREFILL_NODES \
+    DECODE_NODES \
+    RANDOM_RANGE_RATIO
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+set -x
+
+cd "$GITHUB_WORKSPACE/benchmarks/multi_node/vllm_disagg_utils" || exit 1
+
+export TIME_LIMIT="08:00:00"
+export MODEL_PATH=$MODEL_PATH
+export MODEL_NAME=$MODEL_NAME
+export CONTAINER_IMAGE=$IMAGE
+
+if [[ "${PREFILL_EP:-1}" -eq 1 ]]; then
+    export PREFILL_ENABLE_EP=false
+else
+    export PREFILL_ENABLE_EP=true
+fi
+
+if [[ "$PREFILL_DP_ATTN" == "true" ]]; then
+    export PREFILL_ENABLE_DP=true
+else
+    export PREFILL_ENABLE_DP=false
+fi
+
+if [[ "${DECODE_EP:-1}" -eq 1 ]]; then
+    export DECODE_ENABLE_EP=false
+else
+    export DECODE_ENABLE_EP=true
+fi
+
+if [[ "$DECODE_DP_ATTN" == "true" ]]; then
+    export DECODE_ENABLE_DP=true
+else
+    export DECODE_ENABLE_DP=false
+fi
+
+JOB_ID=$(bash ./submit.sh $PREFILL_NODES \
+    $PREFILL_NUM_WORKERS \
+    $DECODE_NODES \
+    $DECODE_NUM_WORKERS \
+    $ISL $OSL "${CONC_LIST// /x}" inf \
+    ${PREFILL_ENABLE_EP} ${PREFILL_ENABLE_DP} \
+    ${DECODE_ENABLE_EP} ${DECODE_ENABLE_DP} \
+    ${PREFILL_TP} ${DECODE_TP} \
+    ${RANDOM_RANGE_RATIO} \
+    "${NODELIST:-}")
+
+if [[ $? -ne 0 ]]; then
+    echo "Failed to submit job" >&2
+    exit 1
+fi
+
+echo "$JOB_ID"

benchmarks/multi_node/vllm_disagg_utils/bench.sh

@@ -0,0 +1,75 @@
+#!/bin/bash
+# vLLM Disaggregated Benchmark Runner
+#
+# Produces JSON result files via benchmark_serving.py (same as SGLang bench.sh)
+# so that the CI pipeline can collect and process results.
+#
+# Usage: bash bench.sh <n_prefill> <n_decode> <prefill_gpus> <decode_gpus> \
+#            <model_dir> <model_name> <log_path> <isl> <osl> \
+#            <concurrency_list> <req_rate> <random_range_ratio> <num_prompts_multiplier>
+
+n_prefill=$1
+n_decode=$2
+prefill_gpus=$3
+decode_gpus=$4
+model_path=$5
+model_name=$6
+MODEL_PATH="${MODEL_PATH:-${model_path}/${model_name}}"
+log_path=$7
+
+chosen_isl=${8:-1024}
+chosen_osl=${9:-1024}
+concurrency_list=${10:-"512x1"}
+chosen_req_rate=${11:-inf}
+random_range_ratio=${12:-0.8}
+num_prompts_multiplier=${13:-10}
+
+IFS='x' read -r -a chosen_concurrencies <<< "$concurrency_list"
+
+ROUTER_PORT="${ROUTER_PORT:-30000}"
+
+echo "Config ${chosen_isl}; ${chosen_osl}; ${chosen_concurrencies[0]}; ${chosen_req_rate}"
+
+profile_folder="${log_path}/vllm_isl_${chosen_isl}_osl_${chosen_osl}"
+mkdir -p "$profile_folder"
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+REPO_ROOT="$(cd "$(dirname "$0")/../../.." && pwd)"
+
+for max_concurrency in "${chosen_concurrencies[@]}"; do
+
+    export_file="${profile_folder}/concurrency_${max_concurrency}_req_rate_${chosen_req_rate}_gpus_$((prefill_gpus+decode_gpus))_ctx_${prefill_gpus}_gen_${decode_gpus}"
+
+    num_prompts=$(( max_concurrency * num_prompts_multiplier ))
+    if [[ "$num_prompts" -lt 16 ]]; then
+        num_prompts=16
+    fi
+
+    echo "profile_folder: $profile_folder"
+    echo "max_concurrency: $max_concurrency"
+    echo "chosen_req_rate: $chosen_req_rate"
+    echo "MODEL_PATH: $MODEL_PATH"
+    echo "ROUTER_PORT: $ROUTER_PORT"
+    echo "chosen_isl: $chosen_isl"
+    echo "chosen_osl: $chosen_osl"
+    echo "num_prompts: $num_prompts"
+    echo "export_file: $export_file"
+
+    run_benchmark_serving \
+        --bench-serving-dir "$REPO_ROOT" \
+        --model "$MODEL_PATH" \
+        --port "$ROUTER_PORT" \
+        --backend openai \
+        --input-len "$chosen_isl" \
+        --output-len "$chosen_osl" \
+        --random-range-ratio "$random_range_ratio" \
+        --num-prompts "$num_prompts" \
+        --max-concurrency "$max_concurrency" \
+        --result-filename "$export_file" \
+        --result-dir /workspace/
+
+    echo "-----------------------------------------"
+    echo "[BENCH] Cooldown: waiting 10s for idle KV block reaper..."
+    sleep 10
+done