# Deepseek-V3-0324

Deepseek-V3-0324 bf16四机部署步骤

## Table of Contents

[TOC]

## 1、环境准备

每个节点准备环境

```shell
docker pull  harbor.sourcefind.cn:5443/dcu/admin/base/vllm:0.9.2-ubuntu22.04-dtk25.04.2-py3.10
docker run --shm-size 500g --network=host --name=limeng_test2 --privileged --device=/dev/kfd --device=/dev/dri --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v ~/:/workspace/ -v /public/opendas/DL_DATA/llm-models/:/home/models:ro -v /opt/hyhal:/opt/hyhal:ro -it harbor.sourcefind.cn:5443/dcu/admin/base/vllm:0.9.2-ubuntu22.04-dtk25.04.2-py3.10 bash
```

## 2、导入环境变量

每个节点环境变量请添加到 ~/.bashrc 后重启容器

```shell
export ALLREDUCE_STREAM_WITH_COMPUTE=1
#BW集群的VLLM_HOST_IP需要设置为ib网卡对应的IP
export VLLM_HOST_IP=$(hostname -I | awk '{print $2}') 
echo $VLLM_HOST_IP
export NCCL_SOCKET_IFNAME=ib0
export GLOO_SOCKET_IFNAME=ib0
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
#这里用ibstat看看哪些网卡是active
export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
export NCCL_MIN_NCHANNELS=16
export NCCL_MAX_NCHANNELS=16
export NCCL_NET_GDR_READ=1
export VLLM_RPC_TIMEOUT=1800000
export NCCL_TOPO_FILE="/workspace/limeng/topo-input.xml"

unset NCCL_ALGO
export NCCL_NET_GDR_LEVEL=7
export NCCL_SDMA_COPY_ENABLE=0

export VLLM_USE_OPT_ZEROS=1
```

## 3、启动ray集群

```shell
主节点执行 
ray start --head --node-ip-address=主节点ip --port=6688 --num-gpus=8 --num-cpus=32

子节点依次执行
ray start --address=主节点ip:6688 --node-ip-address=子节点ip --num-gpus=8 --num-cpus=32
ray start --address=主节点ip:6688 --node-ip-address=子节点ip --num-gpus=8 --num-cpus=32
ray start --address=主节点ip:6688 --node-ip-address=子节点ip --num-gpus=8 --num-cpus=32
```

## 4、主节点启动服务

```shell

model_path=/home/models/DeepSeek-V3-0324-bf16
model=${model_path##*/}
data_type="bfloat16"
tp=32
port=8899
gpu_memory=0.9

#日期目录
log_date=$(date "+%Y-%m-%d")
time=$(date "+%Y-%m-%d-%H-%M-%S")
log_dir="bw1000_${model}/${log_date}"
mkdir -p "${log_dir}"

vllm serve ${model_path} \
 --trust-remote-code \
 --distributed-executor-backend ray \
 --dtype $data_type \
 --tensor-parallel-size $tp \
 --gpu-memory-utilization $gpu_memory \
 --disable-cascade-attn \
 --host 0.0.0.0 \
 --port $port \
 --max-model-len 40960 \
 --max-seq-len-to-capture 40960 \
 --max-num-batched-tokens 40960 \
 --disable-log-requests \
 --max-num-seqs 1024 \
 --block-size 64 \
 --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 3}' \
 --enable-chunked-prefill \
 --enable-prefix-caching \
 2>&1 | tee "${log_dir}/serve_${time}.log"
 
```

在线推理

```Shell
python benchmark_serving.py --model /home/models/DeepSeek-V3-0324-bf16 --dataset-name random --trust-remote-code --random-input-len 12000 --random-output-len 550 --port 8899 --ignore-eos --max-concurrency 6 --num-prompts 12
```