Atlas 800I A2 双机直连部署DeepSeek-R1-w8a8

一、环境信息

1.1、硬件信息

Atlas 800I A2 * 2

1.2、环境信息

操作系统：openEuler 22.03 LTS
NPU驱动：Ascend-hdk-910b-npu-driver 24.1.0 linux-aarch64.run
NPU固件：Ascend-hdk-910b-npu-firware 7.5.0.3.220.run
MindIE镜像：2.0.T3-800I-A2-py311-openeuler24.03-lts

1.3、组网信息

• 一台服务器的NPU-0口对应另一个服务器的NPU-0，然后剩下的依次使用光纤连接；
• NPU侧IP最好在同一个网段里。

二、安装驱动固件

2.1、创建运行用户和所属组

groupadd HwHiAiUser
useradd -g HwHiAiUser -d /home/HwHiAiUser -m HwHiAiUser -s /bin/bash

2.2、安装驱动（根据提示选择是否重启）

./Ascend-hdk-910b-npu-driver_24.1.0_linux-aarch64.run --full --install-for-all

2.3、安装固件（根据提示选择是否重启）

./Ascend-hdk-910b-npu-firmware_7.5.0.3.220.run --full

三、配置参数面（NPU）网络

3.1、配置IP和掩码

# 配置master节点上NPU的IP和掩码
hccn_tool -i 0 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 1 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 2 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 3 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 4 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 5 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 6 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 7 -ip -s address 修改成自己规划的IP netmask 255.255.255.0# 配置node节点上的NPU的IP和掩码
hccn_tool -i 0 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 1 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 2 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 3 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 4 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 5 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 6 -ip -s address 修改成自己规划的IP netmask 255.255.255.0 
hccn_tool -i 7 -ip -s address 修改成自己规划的IP netmask 255.255.255.0

3.2、配置检测IP

检测IP配置成对端NPU口的IP，比如说机器1的IP是： 192.168.10.11，机器2的IP是：192.168.10.19。那机器1的检测IP就是：192.168.10.19，机器2的检测IP就是：192.168.10.11。

# 配置master节点上NPU的侦测IP
hccn_tool -i 0 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 1 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 2 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 3 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 4 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 5 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 6 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 7 -netdetect -s address 修改成自己规划的IP# 配置master节点上NPU的侦测IP
hccn_tool -i 0 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 1 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 2 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 3 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 4 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 5 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 6 -netdetect -s address 修改成自己规划的IP
hccn_tool -i 7 -netdetect -s address 修改成自己规划的IP

3.3、关闭TLS

for i in {0..7}; do hccn_tool -i $i -tls -s enable 0 ; done

3.4、检测配置是否正确

1.1、检查物理链接
for i in {0..7}; do hccn_tool -i $i -lldp -g | grep Ifname; done 1.2、检查链接情况
for i in {0..7}; do hccn_tool -i $i -link -g ; done1.3、检查网络健康情况
for i in {0..7}; do hccn_tool -i $i -net_health -g ; done1.4、查看侦测ip的配置是否正确
for i in {0..7}; do hccn_tool -i $i -netdetect -g ; done1.5、获取每张卡的ip地址
for i in {0..7};do hccn_tool -i $i -ip -g; done1.6、检测NPU底层TLS行为一致性，每台机器需要是一样的值，建议全0
for i in {0..7}; do hccn_tool -i $i -tls -g ; done | grep switch1.7、NPU底层tls校验行为置0操作 
for i in {0..7};do hccn_tool -i $i -tls -s enable 0;done1.8、检测机器间NPU互联情况
hccn_tool -i NPU节点 -ping -g address 另一台机器NPU的IP

四、编辑rank_table_file.json文件

{"server_count": "2","server_list": [{"device": [{"device_id": "0","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "0"},{"device_id": "1","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "1"},{"device_id": "2","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "2"},{"device_id": "3","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "3"},{"device_id": "4","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "4"},{"device_id": "5","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "5"},{"device_id": "6","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "6"},{"device_id": "7","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "7"}],"server_id": "master-业务IP","container_ip": "master-业务IP"},{"device": [{"device_id": "0","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "8"},{"device_id": "1","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "9"},{"device_id": "2","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "10"},{"device_id": "3","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "11"},{"device_id": "4","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "12"},{"device_id": "5","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "13"},{"device_id": "6","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "14"},{"device_id": "7","device_ip": "参数面IP-修改成自己规划的IP","rank_id": "15"}],"server_id": "node-业务IP","container_ip": "node-业务IP"}],"status": "completed","version": "1.0"
}

五、启动容器

docker run -itd --privileged  --name=容器名称 --net=host \--shm-size 500g \--device=/dev/davinci0 \--device=/dev/davinci1 \--device=/dev/davinci2 \--device=/dev/davinci3 \--device=/dev/davinci4 \--device=/dev/davinci5 \--device=/dev/davinci6 \--device=/dev/davinci7 \--device=/dev/davinci_manager \--device=/dev/hisi_hdc \--device /dev/devmm_svm \-v /usr/local/Ascend/driver:/usr/local/Ascend/driver \-v /usr/local/Ascend/firmware:/usr/local/Ascend/firmware \-v /usr/local/sbin/npu-smi:/usr/local/sbin/npu-smi \-v /usr/local/sbin:/usr/local/sbin \-v /etc/hccn.conf:/etc/hccn.conf \-v 物理机权重路径：容器权重路径 \容器镜像名称或ID \bash

六、设置容器环境变量

vim ~/.bashrc# 添加如下配置：
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/atb-models/set_env.sh
source /usr/local/Ascend/mindie/set_env.sh
export ATB_LLM_HCCL_ENABLE=1
export ATB_LLM_COMM_BACKEND="hccl"
export HCCL_CONNECT_TIMEOUT=7200
export WORLD_SIZE=32
export HCCL_EXEC_TIMEOUT=0
export RANKTABLEFILE=/opt/rank_table_file.json
export MIES_CONTAINER_IP= {容器ip地址}
export MINDIE_LOG_STDOUT = 1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export ATB_WORKSPACE_MEM_ALLOC_ALG_TYPE=3
export ATB_WORKSPACE_MEM_ALLOC_GLOBAL=1
export OMP_NUM_THREADS=10
export NPU_MEMORY_FRACTION=0.96
export HCCL_DETERMINISTIC=false
export ATB_LLM_BENCHMARK_ENABLE=1
unset HCCL_OP_EXPANSION_MODE
export HCCL_EXEC_TIMEOUT=0
export TASK_QUEUE_ENABLE=2
export CPU_AFFINITY_CONF=1
export MINDIE_CHECK_INPUTFILES_PERMISSIONS=0
unset HCCL_RDMA_PCIE_DIRECT_POST_NOSTRICT
export MINDIE_LOG_TO_STDOUT="benchmark:1; client:1"

七、拉起服务

# 修改MindIE配置文件
cd /usr/local/Ascend/mindie/latest/mindie-service/
vim conf/config.json# 修改如下配置：
"ipAddress" : "改成主节点IP""managementIpAddress" : "改成主节点IP"
"httpsEnabled" : false
"multiNodesInferEnabled" : true
"interCommTLSEnabled" : false
"interNodeTLSEnabled" : false
"npudeviceIds" : [[0,1,2,3,4,5,6,7]]
"modelName" : "模型名称"
"modelWeightPath" : "权重路径"
"worldSize":8# 配置文件权限
chmod -R 640 {rank_table_file.json路径}
chmod -R 750 {/path-to-weights}# 拉起服务
nohup ./bin/mindieservice_daemon > output.log 2>&1 &# 停止服务
ps -ef | grep mindieservice
pkill -9 mindieservice

八、测试

curl http://IP:1025/v1/chat/completions -d '{"model": "改为mindie配置文件中设置的名称","messages": [{"role": "user", "content": "请告诉我关于人工智能的一些信息。"}],"stream": true}'