Kimi 模型部署

Kimi-K2.5/2.6 模型部署

本章节以 Kimi-K2.5-FP8 部署为例，采取的是最小部署方案 3P4D，即3台机器作为Prefill，4台机器为Decode。支持64K及以上长输入，需增加 Decode 机器数。注意 384（expert数）需要能被 Decode 机器数整除。该部署方式同样适用于 Kimi-K2.6。

• 脚本涉及的参数说明和一键运行说明见 快速开始。
• 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
• 如要开启投机采样，建议使用 4P6D 部署方案，参考 /sgl-workspace/kimi-scripts 目录下的 decode_server_mtp.sh 和 run_mtp.sh 脚本。

信息

启动 SGLang 服务前需要切换python虚拟环境，执行命令：workon kimi

Prefill启动

prefill启动脚本 prefill_server.sh

#!/bin/bash

source /root/.local/bin/uv-virtualenvwrapper.sh
workon kimi

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

export PYTHONPATH=/sgl-workspace/sglang-kimi/python:$PYTHONPATH

export MATE_FORCE_JIT=1
export TP_SOCKET_IFNAME=bond0
export GLOO_SOCKET_IFNAME=bond0
export MCCL_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 SGL_USE_DEEPGEMM_BMM=0
export MC_TE_METRIC=true
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=6000
export SGLANG_PP_LAYER_PARTITION="20,20,21"
export SGLANG_DISAGGREGATION_QUEUE_SIZE=8
export SGLANG_DISAGGREGATION_THREAD_POOL_SIZE=16
export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600
export SGLANG_DISAGGREGATION_MAPPING_IB_DEVICE_TO_GPU=1
export SGLANG_SET_CPU_AFFINITY=1
export SGLANG_ENABLE_TORCH_INFERENCE_MODE=true
export SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1
export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:/usr/local/musa/lib:$LD_LIBRARY_PATH
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 SGL_DEEP_GEMM_BLOCK_M=128
export MC_IB_TC=136
export MC_IB_PCIE_RELAXED_ORDERING=1

export SGLANG_TORCH_PROFILER_DIR=$SCRIPT_DIR/traces
export SGLANG_EXPERT_DISTRIBUTION_RECORDER_DIR=$SCRIPT_DIR/traces
# DeepEP
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=200
export MUSA_DEVICE_PAGE_SIZE=0x1000
export MCCL_CHECK_POINTERS=0

export SGLANG_USE_MTT=1
export SGLANG_SBO_COMBINE_SHARED=1
export SGLANG_DEEPEP_USE_MUSA_ACE=1

NODE_RANK=$1
PREFILL_IP="${2}:5303"
LOG_DIR=$3
MODEL_PATH=$4
HOST_IP=$5
NNODES=$6

SGLANG_PORT=20133
DEEP_EP_CONFIG=$SCRIPT_DIR/deepep.config

mkdir -p "${LOG_DIR}"


# 获取当前环境的 RDMA 设备名列表，在启动时作为 disaggregation-ib-device 参数值
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 "$port/link_layer" 2>/dev/null)
        # 状态为 Up
        state=$(cat "$port/state" 2>/dev/null)
        # 只要是 200G/400G 且状态为 Up（IB 或 RoCE 都收）
        if [[ "$rate" == *"200 Gb"* || "$rate" == *"400 Gb"* ]] && [[ "$state" == "4"* ]]; 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"


nohup python3 -m sglang.launch_server \
            --model-path $MODEL_PATH \
	    --served-model-name kimi-k2.5-fp8 \
	    --enable-single-batch-overlap \
            --trust-remote-code \
	    --disable-overlap-schedule \
            --host 0.0.0.0 \
            --port $SGLANG_PORT \
            --disable-cuda-graph \
	    --attention-backend fa3 \
	    --prefill-attention-backend fa3 \
	    --sampling-backend flashinfer \
            --disaggregation-mode prefill \
            --dist-init-addr $PREFILL_IP \
	    --enable-cache-report \
 	    --chunked-prefill-size 65536 \
	    --max-prefill-tokens 81920 \
            --nnodes $NNODES  \
            --node-rank $NODE_RANK \
	    --pp-size $NNODES \
            --dp-size 8 \
	    --ep-size 8 \
	    --tp-size 8 \
	    --enable-nan-detection \
	    --context-length 240000 \
	    --moe-a2a-backend deepep \
	    --deepep-mode normal \
	    --moe-dense-tp-size 1 \
	    --log-level debug \
	    --enable-dp-attention \
	    --enable-dp-lm-head \
	    --load-balance-method round_robin \
	    --deepep-config "$DEEP_EP_CONFIG" \
	    --max-running-requests 128 \
	    --enable-cache-report \
            --mem-fraction-static 0.8 \
            --disable-radix-cache \
	    --schedule-conservativeness 1 \
	    --disable-shared-experts-fusion \
	    --tool-call-parser kimi_k2 \
	    --reasoning-parser kimi_k2 \
	    --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

source /root/.local/bin/uv-virtualenvwrapper.sh
workon kimi

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

export PYTHONPATH=/sgl-workspace/sglang-kimi/python/:$PYTHONPATH

export MATE_FORCE_JIT=1
export GLOO_SOCKET_IFNAME=bond0
export TP_SOCKET_IFNAME=bond0
export MCCL_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 SGL_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_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 MC_IB_TC=136
export MC_IB_PCIE_RELAXED_ORDERING=1
export SGL_DISAGGREGATION_MAPPING_IB_DEVICE_TO_GPU=1
export MCCL_CHECK_POINTERS=0
export SGLANG_ENABLE_ATTN_TP_PYNCCL=1
export MUSA_DEVICE_PAGE_SIZE=0x1000

export SGLANG_TORCH_PROFILER_DIR=$SCRIPT_DIR/traces
export SGLANG_EXPERT_DISTRIBUTION_RECORDER_DIR=$SCRIPT_DIR/traces
# DeepEP
export SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=200

export SGLANG_SBO_COMBINE_SHARED=1
export SGLANG_USE_MTT=1

NODE_RANK=$1
DECODER_IP="${2}:5403"
LOG_DIR=$3
MODEL_PATH=$4
HOST_IP=$5
NNODES=$6

SGLANG_PORT=20143

mkdir -p "${LOG_DIR}"

# 获取当前环境的 RDMA 设备名列表，在启动时作为 disaggregation-ib-device 参数值
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 "$port/link_layer" 2>/dev/null)
        # 状态为 Up
        state=$(cat "$port/state" 2>/dev/null)
        # 只要是 200G/400G 且状态为 Up（IB 或 RoCE 都收）
        if [[ "$rate" == *"200 Gb"* || "$rate" == *"400 Gb"* ]] && [[ "$state" == "4"* ]]; 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"


nohup python3 -m sglang.launch_server \
        --model $MODEL_PATH \
        --trust-remote-code \
        --disable-radix-cache \
        --allow-auto-truncate \
        --max-total-tokens 600000 \
        --tp-size $(($NNODES * 8)) \
        --ep-size $(($NNODES * 8)) \
        --dp-size $(($NNODES * 8)) \
        --pp-size 1 \
        --mem-fraction-static 0.8 \
        --port $SGLANG_PORT \
        --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 \
        --dist-init-addr $DECODER_IP \
	      --cuda-graph-bs 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 \
        --nnodes $NNODES \
        --node-rank $NODE_RANK \
	      --max-running-requests 1920 \
        --host 0.0.0.0 \
	      --enable-nan-detection \
	      --prefill-round-robin-balance \
        --disaggregation-mode decode \
        --disable-shared-experts-fusion \
        --enable-single-batch-overlap \
        --load-balance-method round_robin \
        --mm-attention-backend triton_attn \
        --disaggregation-ib-device $SGLANG_IB_DEVICES \
	      --tool-call-parser kimi_k2 \
	      --reasoning-parser kimi_k2  > "${LOG_DIR}/D${NODE_RANK}_${HOST_IP}_$(date '+%Y%m%d_%H%M%S').log" 2>&1 &

Router启动

Router 启动脚本 router.sh

#!/bin/bash

source /root/.local/bin/uv-virtualenvwrapper.sh
workon kimi

export PYTHONPATH=/sgl-workspace/sglang-kimi/python:$PYTHONPATH

PREFILL_IP="http://${1}:20133"
DECODE_IP="http://${2}:20143"
LOG_DIR=$3

ulimit -n 65535
nohup python -m sglang_router.launch_router \
    --pd-disaggregation \
    --prefill $PREFILL_IP \
    --decode $DECODE_IP \
    --host 0.0.0.0 \
    --request-timeout-secs 7200 \
    --port 30000 > "${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 ))
port=62216

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}"
  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"
set +u

echo -e "\033[32mWORK_HOME: $WORK_HOME\033[0m"
echo -e "\033[32mMODEL_PATH: $MODEL_PATH\033[0m"
echo -e "\033[32mPREFILL_CRIPT_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 $port -f -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 $port -f -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 $port -f -n $PREFILL_MASTER_IP $cmd
echo -e "Main log file: \033[34m$LOG_DIR\033[0m"

验证 API

在运行 Router 的机器上（若 --net host，容器内访问即可），验证是否有正常输出

curl http://127.0.0.1:30000/v1/chat/completions  \
    -H "Content-Type: application/json"          \
    -d '{     
        "model": "kimi-k2.5-fp8",  
        "messages": [{"role": "user", "content": "你好，SGLang！"}],  
        "max_tokens": 100,      
        "temperature": 0.7    
  }'

预期返回 JSON，choices[0].message.content 为模型回复（示例与线上模型权重、路由配置有关）。

---