跳到主要内容

Deepseek 模型部署

本章节包含 Deepseek V3.2 和 Deepseek-R1 部署脚本。

  • 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
  • Deepseek-R1 使用的 EPLP 文件下载地址：EPLP。该策略只做参考，MOE 模型的 EPLB 策略建议用户根据数据集或线上流量自行确定。
信息

启动 SGLang 服务前切换python虚拟环境到sglang-0.5.6 workon sglang-0.5.6 或 source ~/.virtualenvs/sglang-0.5.6/bin/activate

清理triton缓存：rm -rf ~/.triton/cache/ && rm -r /tmp/*

Deepseek V3.2​

该示例启动方式为2P4D，请准备6台机器。最小部署方式为2机混布（如 TP8 PP2）。

Prefill启动脚本 prefill_server.sh

#!/bin/bash
source ~/.virtualenvs/sglang-0.5.6/bin/activate

export SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_IDLE=0
export TP_SOCKET_IFNAME=bond1
export GLOO_SOCKET_IFNAME=bond1
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="32,29"
#export SGLANG_PP_LAYER_PARTITION="31,30"
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 LD_LIBRARY_PATH=/usr/local/musa/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu/:${LD_LIBRARY_PATH}
export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:/usr/local/musa/lib:$LD_LIBRARY_PATH


export SGLANG_USE_MUSA_FUSED_KERNEL=1
export VLLM_CONFIGURE_LOGGING=0
export VLLM_PATCH_MUSA_CUSTOM_OPS=1

export MATE_FORCE_JIT=1
export SGLANG_DEEPEP_BF16_DISPATCH=0

export NVSHMEM_IB_TRAFFIC_CLASS=136
#export MC_IB_TC=136

#export SGLANG_USE_CUDA_ROTARY_EMBEDDING=1

# export SGLANG_USE_MTT=1

#export MC_ENABLE_DEST_DEVICE_AFFINITY=1
export SGLANG_DISAGGREGATION_MAPPING_IB_DEVICE_TO_GPU=1
export SGLANG_DEEP_GEMM_BLOCK_M=256

export MC_IB_TC=136
export MC_IB_PCIE_RELAXED_ORDERING=1

# 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"
# 获取 RDMA 设备名列表结束

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

PROFILER_DIR_BASE="$PWD"
export SGLANG_TORCH_PROFILER_DIR="${PROFILER_DIR_BASE}/traces/sglang_prefill_profiling_p${NODE_RANK}"

mkdir -p "${LOG_DIR}"

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
DEEP_EP_CONFIG="${SCRIPT_DIR}/deepep.config"

export SGLANG_EXPERT_DISTRIBUTION_RECORDER_DIR=/home/dist/qn/scripts/ning.qiao/dsv32

export LD_LIBRARY_PATH=/usr/local/lib/:/usr/local/mtshmem/lib/:/usr/local/musa/lib:/usr/local/musa/lib:/usr/local/openmpi/lib:/usr/local/musa/mudnn/lib:/usr/lib/x86_64-linux-gnu:ldconfig

#echo -e "\033[32mLOG_DIR: ${LOG_DIR}\033[0m"
nohup python3 -m sglang.launch_server \
            --model-path $MODEL_PATH \
            --served-model-name deepseek-v3.2\
            --trust-remote-code \
            --disable-cuda-graph \
            --disable-overlap-schedule \
            --host 0.0.0.0 \
            --port 20133 \
            --attention-backend nsa \
            --nsa-prefill-backend tilelang \
            --nsa-decode-backend tilelang \
            --sampling-backend flashinfer \
            --disaggregation-mode prefill \
            --dist-init-addr $PREFILL_IP \
            --chunked-prefill-size -1 \
            --max-prefill-tokens 4096 \
            --nnodes 2  \
            --node-rank $NODE_RANK \
            --tensor-parallel-size 8 \
            --ep-size 8 \
            --dp-size 8 \
	          --context-length 6000 \
            --pp-size 2 \
            --moe-a2a-backend deepep \
            --deepep-mode normal \
            --moe-dense-tp-size 1 \
            --enable-dp-attention \
            --enable-dp-lm-head \
            --load-balance-method round_robin \
            --deepep-config "$DEEP_EP_CONFIG" \
            --max-running-requests 64 \
            --enable-cache-report \
            --mem-fraction-static 0.8 \
	          --disable-radix-cache \
            --schedule-conservativeness 1 \
	          --disaggregation-ib-device $SGLANG_IB_DEVICES \
            --disable-shared-experts-fusion > "${LOG_DIR}/P${NODE_RANK}_${HOST_IP}_$(date '+%Y%m%d_%H%M%S').log" 2>&1 &

Decode启动脚本 decode_server.sh

source ~/.virtualenvs/sglang-0.5.6/bin/activate

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_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=16
export SGLANG_USE_DEEPGEMM_BMM=0
export MC_TE_METRIC=true
export SGLANG_DISAGGREGATION_WAITING_TIMEOUT=600
export SGLANG_TORCH_PROFILER_DIR=/home/dist/qn/scripts/ning.qiao/dsv32

export NVSHMEM_IB_ENABLE_SHUFFLE=0 
export NVSHMEM_ENABLE_NIC_PE_MAPPING=1 
export NVSHMEM_HCA_PE_MAPPING=mlx5_bond_2:1:2,mlx5_bond_3:1:2,mlx5_bond_4:1:2,mlx5_bond_5:1:2
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 MATE_FORCE_JIT=1

export SGLANG_DEEPEP_LOW_LATENCY_USE_NVLINK=1
#export SGLANG_USE_CUDA_ROTARY_EMBEDDING=1

#export LD_LIBRARY_PATH=/usr/local/musa/lib:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu/:${LD_LIBRARY_PATH}
export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:/usr/local/musa/lib:$LD_LIBRARY_PATH
export SGLANG_DEEPEP_BF16_DISPATCH=0

export MC_IB_TC=136
export MC_IB_PCIE_RELAXED_ORDERING=1

export SGLANG_USE_MTT=0

export NVSHMEM_IB_TRAFFIC_CLASS=136

#export MC_ENABLE_DEST_DEVICE_AFFINITY=1
# export SGL_DISAGGREGATION_MAPPING_IB_DEVICE_TO_GPU=1
export SGLANG_DISAGGREGATION_MAPPING_IB_DEVICE_TO_GPU=1
export MUSA_DEVICE_PAGE_SIZE=0x1000
#export MUSA_MANAGED_FORCE_DEVICE_ALLOC=1
export NVSHMEM_IBGDA_NIC_HANDLER=gpu
export NVSHMEM_IBGDA_FORCE_NIC_BUF_MEMTYPE=gpumem

# 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"
# 获取 RDMA 设备名列表结束

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

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

export LD_LIBRARY_PATH=/usr/local/lib/:/usr/local/mtshmem/lib/:/usr/local/musa/lib:/usr/local/openmpi/lib:/usr/local/musa/mudnn/lib:/usr/lib/x86_64-linux-gnu:ldconfig
nohup python3 -m sglang.launch_server \
        --model $MODEL_PATH \
        --trust-remote-code \
        --disable-cuda-graph-padding \
        --ep-size 32 \
        --tp-size 32 \
        --dp-size 32 \
        --cuda-graph-bs $(seq 1 16) \
        --mem-fraction-static 0.76 \
        --speculative-algorithm EAGLE \
        --speculative-num-steps 3 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 4 \
        --enable-dp-lm-head \
        --enable-dp-attention \
        --port 20143 \
        --moe-a2a-backend deepep \
        --deepep-mode low_latency \
        --dist-init-addr $DECODER_IP \
        --nnodes 4 \
        --node-rank $NODE_RANK \
        --max-running-requests 1920 \
        --attention-backend nsa \
        --nsa-prefill-backend tilelang \
        --nsa-decode-backend tilelang \
        --moe-dense-tp-size 1 \
        --host 0.0.0.0 \
        --sampling-backend flashinfer \
        --disable-shared-experts-fusion \
        --chunked-prefill-size -1 \
        --load-balance-method round_robin \
        --disable-radix-cache \
        --decode-log-interval 1 \
        --enable-single-batch-overlap \
        --disaggregation-ib-device $SGLANG_IB_DEVICES \
        --disaggregation-mode decode \
        --schedule-conservativeness 0.8 \
        --prefill-round-robin-balance > "${LOG_DIR}/D${NODE_RANK}_${HOST_IP}_$(date '+%Y%m%d_%H%M%S').log" 2>&1 &

Router启动脚本 router.sh

source ~/.virtualenvs/sglang-0.5.6/bin/activate

export PATH=/root/.virtualenvs/sglang-0.5.6/bin/:$PATH

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 &

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

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=${PREFILL_SERVER_COUNT:-2}
DECODER_SERVER_COUNT=${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}"
  MODEL_PATH=${MODEL_PATH:-"/data/models/DeepSeek-V3.2/"}
  HOSTFILE=./hostfile
  LOG_DIR=$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="$WORK_HOME/router.sh"
  #ROUTER_SCRIPT_FILE="$WORK_HOME/router_sgl.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 'cd $WORK_HOME;source ~/.virtualenvs/sglang-0.5.6/bin/activate; bash $PREFILL_SCRIPT_FILE $PREFILL_INDEX $PREFILL_MASTER_IP $LOG_DIR $MODEL_PATH $host'"
    #cmd="bash -c 'cd $WORK_HOME; echo $PATH '"
    echo $cmd
    ssh -p 62216 -f -n $host $cmd
  else
    DECODER_INDEX=$((COUNT - PREFILL_SERVER_COUNT))
    cmd="bash -c 'cd $WORK_HOME;source ~/.virtualenvs/sglang-0.5.6/bin/activate; bash $DECODER_SCRIPT_FILE $DECODER_INDEX $DECODER_MASTER_IP $LOG_DIR $MODEL_PATH $host'"
    echo $cmd
    ssh -p 62216 -f -n $host $cmd
  fi
  ((COUNT++))
done
cmd="bash -c 'cd $WORK_HOME;source ~/.virtualenvs/sglang-0.5.6/bin/activate; bash $ROUTER_SCRIPT_FILE $PREFILL_MASTER_IP $DECODER_MASTER_IP $LOG_DIR'"
echo $cmd
ssh -p 62216 -f -n $PREFILL_MASTER_IP $cmd
echo -e "Main log file: \033[34m$LOG_DIR\033[0m"

DeepSeek R1​

该示例启动方式为2P4D，请准备6台机器。最小部署方式为2机混布（如 TP8 PP2）。

Prefill启动脚本 prefill_server.sh

#!/bin/bash
source ~/.virtualenvs/sglang-0.5.6/bin/activate

export TP_SOCKET_IFNAME=bond1
export GLOO_SOCKET_IFNAME=bond1
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="32,29"
#export SGLANG_PP_LAYER_PARTITION="31,30"
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=true

export SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1
export SGLANG_MOE_PADDING=1

#export LD_LIBRARY_PATH=/usr/local/musa/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:/usr/local/musa/lib:$LD_LIBRARY_PATH

export SGLANG_USE_MUSA_FUSED_KERNEL=1
export VLLM_CONFIGURE_LOGGING=0
export VLLM_PATCH_MUSA_CUSTOM_OPS=1

export MATE_FORCE_JIT=1

export NVSHMEM_IB_TRAFFIC_CLASS=136
#export MC_IB_TC=136

#export SGLANG_USE_CUDA_ROTARY_EMBEDDING=1

# export SGLANG_USE_MTT=1

#export MC_ENABLE_DEST_DEVICE_AFFINITY=1
export SGLANG_DISAGGREGATION_MAPPING_IB_DEVICE_TO_GPU=1
export SGL_DEEP_GEMM_BLOCK_M=256

export MC_IB_TC=136
export MC_IB_PCIE_RELAXED_ORDERING=1
export SGLANG_DEEPEP_USE_MUSA_ACE=1

# 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"
# 获取 RDMA 设备名列表结束

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

PROFILER_DIR_BASE="$PWD"
export SGLANG_TORCH_PROFILER_DIR="${PROFILER_DIR_BASE}/sglang_prefill_profiling_P${NODE_RANK}"

mkdir -p "${LOG_DIR}"

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
DEEP_EP_CONFIG="${SCRIPT_DIR}/deepep.config"

EPLB_FILE="${SCRIPT_DIR}/expert_distribution_recorder_ep8_pp2_${NODE_RANK}.pt"

#echo -e "\033[32mLOG_DIR: ${LOG_DIR}\033[0m"
nohup python3 -m sglang.launch_server \
            --model-path $MODEL_PATH \
            --served-model-name deepseek-r1 \
            --trust-remote-code \
            --disable-cuda-graph \
            --disable-overlap-schedule \
            --host 0.0.0.0 \
            --port 20133 \
            --attention-backend fa3 \
            --prefill-attention-backend fa3 \
            --sampling-backend flashinfer \
            --disaggregation-mode prefill \
            --dist-init-addr $PREFILL_IP \
            --chunked-prefill-size -1 \
            --max-prefill-tokens 8192 \
            --nnodes 2  \
            --node-rank $NODE_RANK \
            --tensor-parallel-size 8 \
            --ep-size 8 \
            --dp-size 8 \
	          --context-length 6000 \
            --pp-size 2 \
            --moe-a2a-backend deepep \
            --deepep-mode normal \
            --moe-dense-tp-size 1 \
            --enable-dp-attention \
            --enable-dp-lm-head \
            --load-balance-method round_robin \
            --deepep-config "$DEEP_EP_CONFIG" \
            --max-running-requests 128 \
            --enable-cache-report \
            --mem-fraction-static 0.8 \
            --pp-max-micro-batch-size 1 \
            --disable-radix-cache \
            --enable-single-batch-overlap \
            --schedule-conservativeness 1 \
	          --init-expert-location "$EPLB_FILE" \
            --eplb-rebalance-experts-per-chunk 32 \
	          --ep-dispatch-algorithm dynamic \
	          --enable-eplb \
            --eplb-algorithm deepseek \
            --eplb-rebalance-num-iterations 1000 \
            --eplb-min-rebalancing-utilization-threshold 0.85 \
            --enable-eplb-rebalance-async \
            --eplb-rebalance-layers-per-chunk 1 \
	          --disaggregation-ib-device $SGLANG_IB_DEVICES \
            --disable-shared-experts-fusion > "${LOG_DIR}/P${NODE_RANK}_${HOST_IP}_$(date '+%Y%m%d_%H%M%S').log" 2>&1 &

Decode启动脚本 decode_server.sh

#!/bin/bash
source ~/.virtualenvs/sglang-0.5.6/bin/activate

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_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=160
export SGLANG_USE_DEEPGEMM_BMM=0
export MC_TE_METRIC=true
export SGLANG_DISAGGREGATION_WAITING_TIMEOUT=6000

export NVSHMEM_IB_ENABLE_SHUFFLE=0 
export NVSHMEM_ENABLE_NIC_PE_MAPPING=1 
export NVSHMEM_HCA_PE_MAPPING=mlx5_bond_2:1:2,mlx5_bond_3:1:2,mlx5_bond_4:1:2,mlx5_bond_5:1:2
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 MATE_FORCE_JIT=1

export SGLANG_DEEPEP_LOW_LATENCY_USE_NVLINK=1
#export SGLANG_USE_CUDA_ROTARY_EMBEDDING=1

#export LD_LIBRARY_PATH=/usr/local/musa/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:/usr/local/musa/lib:$LD_LIBRARY_PATH

export MC_IB_TC=136
export MC_IB_PCIE_RELAXED_ORDERING=1

export SGLANG_USE_MTT=0

export NVSHMEM_IB_TRAFFIC_CLASS=136

#export MC_ENABLE_DEST_DEVICE_AFFINITY=1
export SGLANG_DISAGGREGATION_MAPPING_IB_DEVICE_TO_GPU=1
export MUSA_DEVICE_PAGE_SIZE=0x1000
#export MUSA_MANAGED_FORCE_DEVICE_ALLOC=1
export NVSHMEM_IBGDA_NIC_HANDLER=gpu
export NVSHMEM_IBGDA_FORCE_NIC_BUF_MEMTYPE=gpumem

# 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"
# 获取 RDMA 设备名列表结束

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

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

EPLB_FILE="$SCRIPT_DIR/deepseek_r1_eplb_weight-decoder-ep32-mtp-padhi.pt"

PROFILER_DIR_BASE="$PWD"
export SGLANG_TORCH_PROFILER_DIR="${PROFILER_DIR_BASE}/sglang_prefill_profiling_D${NODE_RANK}"

nohup python3 -m sglang.launch_server \
        --model $MODEL_PATH \
        --trust-remote-code \
	--disable-cuda-graph-padding \
	--cuda-graph-bs $(seq 1 54) \
        --tensor-parallel-size 32 \
        --ep-size 32 \
        --dp-size 32 \
        --mem-fraction-static 0.76 \
        --speculative-algorithm EAGLE \
	--speculative-num-steps 1 \
        --speculative-eagle-topk 1 \
        --speculative-num-draft-tokens 2 \
	--enable-dp-lm-head \
        --enable-dp-attention \
        --port 20143 \
        --moe-a2a-backend deepep \
        --deepep-mode low_latency \
        --dist-init-addr $DECODER_IP \
        --nnodes 4 \
        --node-rank $NODE_RANK \
        --max-running-requests 1920 \
        --attention-backend fa3 \
        --moe-dense-tp-size 1 \
        --host 0.0.0.0 \
        --sampling-backend flashinfer \
     	--disable-shared-experts-fusion \
        --chunked-prefill-size -1 \
        --enable-single-batch-overlap \
        --disaggregation-mode decode \
        --disaggregation-ib-device $SGLANG_IB_DEVICES \
	--load-balance-method round_robin \
        --init-expert-location $EPLB_FILE \
        --ep-num-redundant-experts 0 \
	--disable-radix-cache \
        --schedule-conservativeness 0.8 \
        --prefill-round-robin-balance > "${LOG_DIR}/D${NODE_RANK}_${HOST_IP}_$(date '+%Y%m%d_%H%M%S').log" 2>&1 &

Router启动脚本 router.sh

#!/bin/bash
source ~/.virtualenvs/sglang-0.5.6/bin/activate

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 $DECODE_IP \
    --host 0.0.0.0 \
    --port 30000 > "${LOG_DIR}/router_$(date '+%Y%m%d_%H%M%S').log" 2>&1 &

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

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=${PREFILL_SERVER_COUNT:-2}
DECODER_SERVER_COUNT=${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}"
  MODEL_PATH=${MODEL_PATH:-"/data/models/DeepSeek-R1-0528/"}
  HOSTFILE=./hostfile
  LOG_DIR=$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="$WORK_HOME/router.sh"
  #ROUTER_SCRIPT_FILE="$WORK_HOME/router_sgl.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 'cd $WORK_HOME; bash $PREFILL_SCRIPT_FILE $PREFILL_INDEX $PREFILL_MASTER_IP $LOG_DIR $MODEL_PATH $host'"
    #cmd="bash -c 'cd $WORK_HOME; echo $PATH '"
    echo $cmd
    ssh -p 62216 -f -n $host $cmd
  else
    DECODER_INDEX=$((COUNT - PREFILL_SERVER_COUNT))
    cmd="bash -c 'cd $WORK_HOME; bash $DECODER_SCRIPT_FILE $DECODER_INDEX $DECODER_MASTER_IP $LOG_DIR $MODEL_PATH $host'"
    echo $cmd
    ssh -p 62216 -f -n $host $cmd
  fi
  ((COUNT++))
done
cmd="bash -c 'cd $WORK_HOME; bash $ROUTER_SCRIPT_FILE $PREFILL_MASTER_IP $DECODER_MASTER_IP $LOG_DIR'"
echo $cmd
ssh -p 62216 -f -n $PREFILL_MASTER_IP $cmd
echo -e "Main log file: \033[34m$LOG_DIR\033[0m"

版权所有 2024 摩尔线程，本文档受国际版权法保护。

摩尔线程和摩尔线程徽标是摩尔线程智能科技 (北京)有限责任公司的注册商标。

免责声明

本文档提供有关摩尔线程产品的信息。本文档并未授权任何知识产权的许可，并未以明示或暗示，或以禁止反言或其他方式授予任何知识产权许可。