Kimi 模型部署

本章节以 Kimi-K2.5-FP8 部署为例，Kimi-K2.6-FP8 可使用相同方式，更改权重文件路径即可。

• 模型权重下载。

• 参考 环境准备 进行环境检查、启动容器。

• 容器镜像地址参考 版本发布信息-镜像地址。

• PD分离脚本涉及的参数说明和一键运行说明见 快速开始。

模型部署

示例采取的是最小部署方案 P3D4，即3台机器作为Prefill，4台机器为Decode。

• 如果需要修改部署方式，请更新 run.sh 中 PREFILL_SERVER_COUNT 和 DECODE_SERVER_COUNT。

• DeepEP 的配置文件内容参考 DeepEP 配置文件

Prefill启动

prefill启动脚本 prefill_server.sh

#!/bin/bash
export PATH=/root/.virtualenvs/sglang-default/bin/:$PATH

export TP_SOCKET_IFNAME=bond0
export GLOO_SOCKET_IFNAME=bond0
export MUSA_LAUNCH_BLOCKING=0
export MCCL_IB_GID_INDEX=3
export MCCL_NET_SHARED_BUFFERS=0
export MCCL_PROTOS=2
export SGLANG_USE_DEEPGEMM_BMM=0
export MC_TE_METRIC=true
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
export SGLANG_PP_LAYER_PARTITION="20,20,21"
export SGLANG_DISAGGREGATION_QUEUE_SIZE=8
export SGLANG_DISAGGREGATION_THREAD_POOL_SIZE=16
export SGLANG_DISAGGREGATION_MAPPING_IB_DEVICE_TO_GPU=1
export SGLANG_SET_CPU_AFFINITY=1
export SGLANG_ENABLE_TORCH_INFERENCE_MODE=true
export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:/usr/local/musa/lib:$LD_LIBRARY_PATH
export SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1
export SGLANG_MOE_PADDING=1
export SGLANG_USE_MUSA_FUSED_KERNEL=1
export VLLM_CONFIGURE_LOGGING=0
export VLLM_PATCH_MUSA_CUSTOM_OPS=1
# export SGL_DISAGGREGATION_MAPPING_IB_DEVICE_TO_GPU=1
export SGLANG_DEEP_GEMM_BLOCK_M=128
export MC_IB_TC=136
export MC_IB_PCIE_RELAXED_ORDERING=1
export SGLANG_TORCH_PROFILER_DIR=/root/trace_log

# DeepEP
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=200
export MCCL_CHECK_POINTERS=0


# ACE
export SGLANG_DEEPEP_USE_MUSA_ACE=1
# shared experts overlap
export SGLANG_SBO_COMBINE_SHARED=1



NODE_RANK=$1
PREFILL_IP="${2}:12345"
LOG_DIR=$3
MODEL_PATH=$4
HOST_IP=$5
NNODES=$6
echo "NNODES=$NNODES"
IB_DEVS=()

for dev in /sys/class/infiniband/*; do
    ibdev=$(basename "$dev")

    for port in "$dev"/ports/*; do
        # 速率（如 200 Gb/sec）
        rate=$(cat "$port/rate" 2>/dev/null)
        # link_layer: InfiniBand / Ethernet (RoCE)
        link=$(cat "$port/link_layer" 2>/dev/null)
        # 只要是 200G/400G（IB 或 RoCE 都收）
        if [[ "$rate" == *"200 Gb"* || "$rate" == *"400 Gb"* ]]; then
            IB_DEVS+=("$ibdev")
            break
        fi
    done
done
# 排序 + 去重
IB_DEVS_SORTED=$(printf "%s\n" "${IB_DEVS[@]}" | sort -V | uniq)
# 生成 MCCL / SGLang 变量
MCCL_IB_HCA=$(echo "$IB_DEVS_SORTED" | paste -sd, -)
SGLANG_IB_DEVICES="$MCCL_IB_HCA"
echo "IB_DEVICES=$MCCL_IB_HCA"

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
export SGLANG_TORCH_PROFILER_DIR="${SCRIPT_DIR}/traces/sglang_decode_profiling_P${NODE_RANK}"

mkdir -p "${LOG_DIR}"

DEEP_EP_CONFIG="/data/workspace/deepep/deepep.config"

source /root/.virtualenvs/sglang-default/bin/activate
nohup python3 -m sglang.launch_server \
            --model-path $MODEL_PATH \
            --served-model-name base-model \
            --trust-remote-code \
            --disable-cuda-graph \
            --disable-overlap-schedule \
            --host 0.0.0.0 \
            --port 24586 \
            --nnodes $NNODES  \
            --node-rank $NODE_RANK \
            --tp-size 8 \
            --pp-size 3 \
            --dp-size 8  \
            --ep-size 8 \
            --attention-backend fa3 \
            --prefill-attention-backend fa3 \
            --sampling-backend flashinfer \
            --linear-attn-backend flashinfer \
            --disaggregation-mode prefill \
            --moe-runner-backend auto \
            --dist-init-addr $PREFILL_IP \
            --enable-cache-report \
            --enable-nan-detection \
            --moe-a2a-backend deepep \
            --deepep-mode normal \
            --moe-dense-tp-size 1 \
            --enable-dp-attention \
            --enable-dp-lm-head \
            --load-balance-method auto \
            --deepep-config "$DEEP_EP_CONFIG" \
            --max-running-requests 8 \
            --enable-cache-report \
            --mem-fraction-static 0.86 \
            --schedule-conservativeness 1 \
            --disable-shared-experts-fusion \
            --pp-max-micro-batch-size 1 \
            --enable-single-batch-overlap \
            --max-total-tokens 270000 \
            --chunked-prefill-size 65536 \
            --max-prefill-tokens 270000 \
            --context-length 266240 \
            --disaggregation-ib-device $SGLANG_IB_DEVICES > "${LOG_DIR}/P${NODE_RANK}_${HOST_IP}_$(date '+%Y%m%d_%H%M%S').log" 2>&1 &

Decode启动

decoder启动脚本 decode_server.sh

#!/bin/bash
export PATH=/root/.virtualenvs/sglang-default/bin/:$PATH

export GLOO_SOCKET_IFNAME=bond0
export TP_SOCKET_IFNAME=bond0
export MUSA_LAUNCH_BLOCKING=0
export MCCL_IB_GID_INDEX=3
export MCCL_NET_SHARED_BUFFERS=0
export MCCL_PROTOS=2
export SGLANG_USE_DEEPGEMM_BMM=0
export MC_TE_METRIC=true
export SGLANG_DISAGGREGATION_WAITING_TIMEOUT=6000
export SGLANG_DISAGGREGATION_THREAD_POOL_SIZE=16
export SGLANG_DISAGGREGATION_QUEUE_SIZE=8
export SGLANG_DISAGGREGATION_MAPPING_IB_DEVICE_TO_GPU=1
export SGLANG_SET_CPU_AFFINITY=1
export SGLANG_ENABLE_TORCH_INFERENCE_MODE=false
export SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1
export SGLANG_MOE_PADDING=1
export SGLANG_USE_MUSA_FUSED_KERNEL=1
export VLLM_CONFIGURE_LOGGING=0
export VLLM_PATCH_MUSA_CUSTOM_OPS=1
export SGLANG_DEEPEP_LOW_LATENCY_USE_NVLINK=1
export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:/usr/local/musa/lib:$LD_LIBRARY_PATH
export SGLANG_USE_MTT=0
export MCCL_CHECK_POINTERS=0

# DeepEP
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=200
export NVSHMEM_IB_SL=7
export MUSA_DEVICE_PAGE_SIZE=0x1000
export MUSA_MANAGED_FORCE_DEVICE_ALLOC=1
export NVSHMEM_IBGDA_FORCE_NIC_BUF_MEMTYPE=gpumem
export NVSHMEM_IBGDA_NIC_HANDLER=cpu
export NVSHMEM_IBGDA_NUM_PROXY_PER_PE=2

# SBO
export SGLANG_SBO_COMBINE_SHARED=1


NODE_RANK=$1
DECODER_IP="${2}:14343"
LOG_DIR=$3
MODEL_PATH=$4
HOST_IP=$5
NNODES=$6
WORLD_SIZE=$NNODES
echo "NNODES=$NNODES"
IB_DEVS=()

for dev in /sys/class/infiniband/*; do
    ibdev=$(basename "$dev")

    for port in "$dev"/ports/*; do
        # 速率（如 200 Gb/sec）
        rate=$(cat "$port/rate" 2>/dev/null)
        # link_layer: InfiniBand / Ethernet (RoCE)
        link=$(cat "$port/link_layer" 2>/dev/null)
        # 只要是 200G/400G（IB 或 RoCE 都收）
        if [[ "$rate" == *"200 Gb"* || "$rate" == *"400 Gb"* ]]; then
            IB_DEVS+=("$ibdev")
            break
        fi
    done
done
# 排序 + 去重
IB_DEVS_SORTED=$(printf "%s\n" "${IB_DEVS[@]}" | sort -V | uniq)
# 生成 MCCL / SGLang 变量
MCCL_IB_HCA=$(echo "$IB_DEVS_SORTED" | paste -sd, -)
SGLANG_IB_DEVICES="$MCCL_IB_HCA"
echo "IB_DEVICES=$SGLANG_IB_DEVICES"

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"


export SGLANG_TORCH_PROFILER_DIR="${SCRIPT_DIR}/traces/sglang_decode_profiling_D${NODE_RANK}"

source /root/.virtualenvs/sglang-default/bin/activate
nohup python3 -m sglang.launch_server \
        --model $MODEL_PATH \
        --trust-remote-code \
        --host 0.0.0.0 \
        --port 23457 \
        --tp-size $((WORLD_SIZE * 8)) \
        --ep-size $((WORLD_SIZE * 8)) \
        --dp-size $((WORLD_SIZE * 8)) \
        --pp-size 1 \
        --mem-fraction-static 0.88 \
        --moe-a2a-backend deepep \
        --deepep-mode low_latency \
        --moe-dense-tp-size 1 \
        --disable-shared-experts-fusion \
        --enable-dp-lm-head \
        --enable-dp-attention \
        --schedule-conservativeness 0.8 \
        --attention-backend fa3 \
        --sampling-backend flashinfer \
        --linear-attn-backend flashinfer \
        --dist-init-addr $DECODER_IP \
        --nnodes $NNODES \
        --node-rank $NODE_RANK \
        --chunked-prefill-size -1 \
        --enable-nan-detection \
        --prefill-round-robin-balance \
        --disable-shared-experts-fusion \
        --enable-single-batch-overlap \
        --load-balance-method auto \
        --disaggregation-mode decode \
        --disaggregation-ib-device $SGLANG_IB_DEVICES \
        --cuda-graph-bs $(seq 1 2) \
        --max-total-tokens 270000 \
        --context-length 266240 \
        --max-running-requests 32 \
        --dtype bfloat16 \
        --kv-cache-dtype auto \
        --speculative-draft-load-format auto  > "${LOG_DIR}/D${NODE_RANK}_${HOST_IP}_$(date '+%Y%m%d_%H%M%S').log" 2>&1 &


        # 	--speculative-algorithm EAGLE3  \
        # --speculative-num-steps 1 \
        # --speculative-eagle-topk 1 \
        # --speculative-num-draft-tokens 2 \
        # --speculative-draft-model-path /home/dist/models/kimi-k2.5-draft/ \

Router启动

Router 启动脚本 router.sh

PREFILL_IP="http://$1:24586"
DECODE_IP="http://$2:23457"
LOG_DIR=$3

ulimit -n 65535
source /root/.virtualenvs/sglang-default/bin/activate
nohup python -m sglang_router.launch_router \
    --pd-disaggregation \
    --mini-lb \
    --prefill $PREFILL_IP \
    --decode $DECODE_IP \
    --host 0.0.0.0 \
    --request-timeout-secs 7200 \
    --port 31000 > "${LOG_DIR}/router_$(date '+%Y%m%d_%H%M%S').log" 2>&1 &

创建 hostfile 文件

hostfile文件示例

192.168.100.101
192.168.100.102
192.168.100.103
192.168.100.104
192.168.100.105
192.168.100.106
192.168.100.107

一键启动所有服务

一键运行脚本 run.sh

#!/bin/bash

CURRENT_TIME=$(date "+%Y%m%d_%H%M%S")
#echo $CURRENT_TIME
#mkdir -p ./output/$CURRENT_TIME

PREFILL_SERVER_COUNT=${PREFILL_SERVER_COUNT:-3}
DECODER_SERVER_COUNT=${DECODER_SERVER_COUNT:-4}
WORLD_SIZE=$(( (PREFILL_SERVER_COUNT + DECODER_SERVER_COUNT) * 8 ))

echo -e "\033[32mPREFILL_SERVER_COUNT: $PREFILL_SERVER_COUNT, DECODER_SERVER_COUNT: $DECODER_SERVER_COUNT, \033[0m"

set -u
  WORK_HOME="$PWD"
  EXPNAME="P${PREFILL_SERVER_COUNT}_D${DECODER_SERVER_COUNT}_gpus${WORLD_SIZE}"
  SGLANG_CODE_HOME="${SGLANG_CODE_HOME:-/sgl-workspace/sglang}"
  SGLANG_CODE_HOME_ENV="$SGLANG_CODE_HOME"
  MODEL_PATH=${MODEL_PATH:-"/data/models/kimi-k2.5-fp8/"}
  HOSTFILE=./hostfile
  LOG_DIR=$WORK_HOME/output/$EXPNAME/$CURRENT_TIME/
  PREFILL_SCRIPT_FILE="$WORK_HOME/prefill_server.sh"
  DECODER_SCRIPT_FILE="$WORK_HOME/decode_server.sh"
  ROUTER_SCRIPT_FILE="$WORK_HOME/router.sh"
  SSH_PORT=${SSH_PORT:-62216}
set +u

echo -e "\033[32mWORK_HOME: $WORK_HOME\033[0m"
echo -e "\033[32mMODEL_PATH: $MODEL_PATH\033[0m"
echo -e "\033[32mPREFILL_SCRIPT_FILE: $PREFILL_SCRIPT_FILE\033[0m"
echo -e "\033[32mDECODER_SCRIPT_FILE: $DECODER_SCRIPT_FILE\033[0m"

# bash generate_hostfile.sh 
COUNT=0
hostlist=$(grep -v '^#\|^$' $HOSTFILE | awk '{print $1}' | xargs)

# Check if hostlist is empty
if [ -z "$hostlist" ]; then
  echo "Error: hostlist is empty. Please add IP addresses to the hostfile."
  exit 1
fi

PREFILL_MASTER_IP=$(head -n 1 $HOSTFILE)
DECODER_MASTER_IP=$(tail -n $DECODER_SERVER_COUNT $HOSTFILE | head -n1)
echo -e "\033[32mPREFILL_MASTER_IP: $PREFILL_MASTER_IP\033[0m"
echo -e "\033[32mDECODER_MASTER_IP: $DECODER_MASTER_IP\033[0m"


mkdir -p $LOG_DIR
for host in ${hostlist[@]}; do
  echo $host
  if [[ $COUNT -lt $PREFILL_SERVER_COUNT ]]; then
    PREFILL_INDEX=$COUNT
    cmd="bash -c 'bash $PREFILL_SCRIPT_FILE $PREFILL_INDEX $PREFILL_MASTER_IP $LOG_DIR $MODEL_PATH $host $PREFILL_SERVER_COUNT'"
    #cmd="bash -c 'cd $WORK_HOME; echo $PATH '"
    echo $cmd
    ssh -p $SSH_PORT -n $host $cmd
  else
    DECODER_INDEX=$((COUNT - PREFILL_SERVER_COUNT))
    cmd="bash -c 'bash $DECODER_SCRIPT_FILE $DECODER_INDEX $DECODER_MASTER_IP $LOG_DIR $MODEL_PATH $host $DECODER_SERVER_COUNT'"
    echo $cmd
    ssh -p $SSH_PORT -n $host $cmd
  fi

  ((COUNT++))
done
cmd="bash -c 'bash $ROUTER_SCRIPT_FILE $PREFILL_MASTER_IP $DECODER_MASTER_IP $LOG_DIR'"
echo $cmd
ssh -p $SSH_PORT -n $PREFILL_MASTER_IP $cmd
echo -e "Main log file: \033[34m$LOG_DIR\033[0m"

---