【昇腾】单张96G Atlas 300I Duo推理卡MindIE+WebUI方式跑32B大语言模型_20250818
一、Atlas 300I Duo推理卡相关安装步骤
由于显存的瓶颈，48G的Atlas 300I Duo推理卡是没办法跑得起来DeepSeek-R1-Distill-Qwen-32B大语言模型的，这里换了一张96G版本的Atlas 300I Duo推理卡来跑，32B大语言模组除了对显存有要求，对服务器本身的内存条也有要求，在加载的过程中需要较大的内存，这里服务器的内存条内存为128GB
1.1 服务器系统与内核说明
| 服务器 | 系统版本 | 内核版本 | 内存条内存 |
|---|---|---|---|
| S5000C | Kylin V10 | 4.19.90-89.11.v2401.ky10.aarch64 | 128GB |
P.S.服务器安装好系统后先不要执行yum update -y更新，否则内核版本会从4.19.90-89.11升级到4.19.90-89.21，Atlas 300I Duo推理卡的driver包会安装失败
1.2 系统环境说明
本服务器IP地址：192.168.2.71
登录用户：root
新开一个 terminal ，执行以下命令确认是否有探到Atlas 300I Duo推理卡：
lspci |grep Huawei 如有卡，回显信息是：
0000:01:00.0 Processing accelerators: Huawei Technologies Co., Ltd. Device d500 (rev 23)
uname-a回显信息是：
Linux localhost.localdomain 4.19.90-89.11.v2401.ky10.aarch64 #1 SMP Thu Apr 25 18:20:10 CST 2024 aarch64 aarch64 aarch64 GNU/Linux
cat /etc/*release 回显信息是：
Kylin Linux Advanced Server release V10 (Halberd)
DISTRIB_ID=Kylin
DISTRIB_RELEASE=V10
DISTRIB_CODENAME=Halberd
DISTRIB_DESCRIPTION=“Kylin V10”
DISTRIB_KYLIN_RELEASE=V10
DISTRIB_VERSION_TYPE=enterprise
DISTRIB_VERSION_MODE=normal
NAME=“Kylin Linux Advanced Server”
VERSION=“V10 (Halberd)”
ID=“kylin”
VERSION_ID=“V10”
PRETTY_NAME=“Kylin Linux Advanced Server V10 (Halberd)”
ANSI_COLOR=“0;31”
Kylin Linux Advanced Server release V10 (Halberd)
############################################################################################
1.3 准备安装驱动固件
1.3.1 新增HwHiAiUser用户
groupadd HwHiAiUser useradd-g HwHiAiUser -d /home/HwHiAiUser -m HwHiAiUser -s /bin/bash 1.3.2 准备驱动与固件文件并安装
Ascend-hdk-310p-npu-driver_24.1.0.1_linux-aarch64.run
Ascend-hdk-310p-npu-firmware_7.5.0.5.220.run
将下载好安装文件，放到/root/work目录下:
cd /root/work chmod +x * 参考《Atlas 中心推理卡 24.1.0 NPU驱动和固件安装指南 02.pdf》文档“2 物理机安装与卸载”章节中介绍的方法安装驱动与固件
因为Atlas 300I Duo推理卡是新采购回来的卡，本次安装为首次安装场景，需先安装驱动再安装固件
安装driver:
./Ascend-hdk-310p-npu-driver_24.1.0.1_linux-aarch64.run --check ./Ascend-hdk-310p-npu-driver_24.1.0.1_linux-aarch64.run --full安装成功回显信息是：
Driver package installed successfully! The new version takes effect immediately.
安装firmware:
./Ascend-hdk-310p-npu-firmware_7.5.0.5.220.run --check ./Ascend-hdk-310p-npu-firmware_7.5.0.5.220.run --full安装成功回显信息是：
Firmware package installed successfully! Reboot now or after driver installation for the installation/upgrade to take effect.
执行reboot命令重启
如果驱动固件安装正确，执行 npu-smi info命令探到信息如下：
以上驱动固件安装完毕
############################################################################################
二、安装docker
Kylin V10并没有自带docker命令，需自行安装，请参考：https://blog.ZEEKLOG.net/weixin_43273656/article/details/145469516
2.1 查看内核版本
uname -a 回显信息是：
Linux localhost.localdomain 4.19.90-89.11.v2401.ky10.aarch64 #1 SMP Thu Apr 25 18:20:10 CST 2024 aarch64 aarch64 aarch64 GNU/Linux
2.2 查看内核参数
cat /proc/version 回显信息是：
Linux version 4.19.90-89.11.v2401.ky10.aarch64 ([email protected]) (gcc version 7.3.0 (GCC)) #1 SMP Thu Apr 25 18:20:10 CST 2024
2.3 查看系统和内核的详细信息
hostnamectl 回显信息是：
Static hostname: localhost.localdomain
Icon name: computer-server
Chassis: server
Machine ID: 889689ba3a9f48c4985c1519c2d8f553
Boot ID: 24cf07b36d6d4db69befaca323c4be93
Operating System: Kylin Linux Advanced Server V10 (Halberd)
Kernel: Linux 4.19.90-89.11.v2401.ky10.aarch64
Architecture: arm64
总结：需要下载aarch64的官方下载docker离线安装包，这里下载docker-27.2.0.tgz
2.4 将下载好安装文件，放到/root/work目录下，解压安装包
cd /root/work tar-zxvf docker-27.2.0.tgz 2.5 移动 Docker 文件
mv /root/work/docker/* /usr/bin/ 2.6 修改docker.service
vim /usr/lib/systemd/system/docker.service 新增以下内容：
[Unit]
Description=Docker Application Container Engine
Documentation=https://docs.docker.com
After=network-online.target firewalld.service
Wants=network-online.target
[Service]
Type=notify
ExecStart=/usr/bin/dockerd
ExecReload=/bin/kill -s HUP $MAINPID
LimitNOFILE=infinity
LimitNPROC=infinity
TimeoutStartSec=0
Delegate=yes
KillMode=process
Restart=on-failure
StartLimitBurst=3
StartLimitInterval=60s
[Install]
WantedBy=multi-user.target
2.7 修改daemon.json文件
mkdir-p /etc/docker vim /etc/docker/daemon.json 新增以下内容
{
“exec-opts”: [“native.cgroupdriver=systemd”],
“insecure-registries”: [
“http://172.31.192.88:81”,“http://111.51.123.456:2222”
]
}
2.8 运行守护进程，启动 Docker
dockerd 2.9 docker 其他命令介绍
- 启动
systemctl start docker - 查看状态
systemctl status docker - 设置开机自启动
systemctl enable docker reboot重启设备，并完成以下操作
三、安装与部署
3.1 拉取镜像
https://www.hiascend.com/developer/ascendhub/detail/af85b724a7e5469ebd7ea13c3439d48f
切到镜像版本页面，找到1.0.0-300I-Duo-py311-openeuler24.03-lts镜像点击下载，按指引将镜像拉取到服务器
3.1.1.docker login -u cn-south-1@HST3UBLG0X38GM0FMAGK swr.cn-south-1.myhuaweicloud.com
3.1.2.密码[d153e20f53b515e9f388f5bedf341c09b22b573e143c0cf33e1dd1f834535862]
3.1.3.docker pull swr.cn-south-1.myhuaweicloud.com/ascendhub/mindie:1.0.0-300I-Duo-py311-openeuler24.03-lts
拉取镜像完毕以后：
执行docker images
回显信息是：
3.2 新建容器
docker run -it -d --net=host --shm-size=1g
–privileged
–name sakway
-v /usr/local/Ascend/driver:/usr/local/Ascend/driver:ro
-v /usr/local/sbin:/usr/local/sbin:ro
-v /root/work:/root/work:rw
-v /path-to-weights:/path-to-weights:ro
swr.cn-south-1.myhuaweicloud.com/ascendhub/mindie:1.0.0-300I-Duo-py311-openeuler24.03-lts bash
3.3 查询正在运行的docker
[root@localhost work]# docker ps -a
3.4 进容器：
docker exec -it sakway bash
3.5 下载权重模型：
3.5.1 确保进docker以后：
cd /root/work/ 3.5.2 安装modelscope命令：
pip install modelscope --index-url https://mirrors.huaweicloud.com/repository/pypi/simple/ 3.5.3 下载权重：
modelscope download --model deepseek-ai/DeepSeek-R1-Distill-Qwen-32B 将权重移动到/root/work/目录
mv /root/.cache/modelscope/hub/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B /root/work/ 将权重文件放在/root/work/目录以后，把x权限去掉，给添加config.json文件赋750权限
cd /root/work/ chmod750 /root/work/DeepSeek-R1-Distill-Qwen-32B/config.json vim /root/work/DeepSeek-R1-Distill-Qwen-32B/config.json 将"torch_dtype": “bfloat16”,修改为"torch_dtype": “float16”,
vim /usr/local/Ascend/mindie/latest/mindie-service/conf/config.json 有九处要修改，在/usr/local/Ascend/mindie/latest/mindie-service/conf/目录下有修改以后的config.json与原始的config.json_org，具体修改项可对比
{"Version":"1.1.0","LogConfig":{"logLevel":"Info","logFileSize":20,"logFileNum":20,"logPath":"logs/mindservice.log"},"ServerConfig":{"ipAddress":"192.168.2.71","managementIpAddress":"127.0.0.2","port":1040,"managementPort":1041,"metricsPort":1042,"allowAllZeroIpListening":false,"maxLinkNum":1000,"httpsEnabled":false,"fullTextEnabled":false,"tlsCaPath":"security/ca/","tlsCaFile":["ca.pem"],"tlsCert":"security/certs/server.pem","tlsPk":"security/keys/server.key.pem","tlsPkPwd":"security/pass/key_pwd.txt","tlsCrlPath":"security/certs/","tlsCrlFiles":["server_crl.pem"],"managementTlsCaFile":["management_ca.pem"],"managementTlsCert":"security/certs/management/server.pem","managementTlsPk":"security/keys/management/server.key.pem","managementTlsPkPwd":"security/pass/management/key_pwd.txt","managementTlsCrlPath":"security/management/certs/","managementTlsCrlFiles":["server_crl.pem"],"kmcKsfMaster":"tools/pmt/master/ksfa","kmcKsfStandby":"tools/pmt/standby/ksfb","inferMode":"standard","interCommTLSEnabled":true,"interCommPort":1121,"interCommTlsCaPath":"security/grpc/ca/","interCommTlsCaFiles":["ca.pem"],"interCommTlsCert":"security/grpc/certs/server.pem","interCommPk":"security/grpc/keys/server.key.pem","interCommPkPwd":"security/grpc/pass/key_pwd.txt","interCommTlsCrlPath":"security/grpc/certs/","interCommTlsCrlFiles":["server_crl.pem"],"openAiSupport":"vllm"},"BackendConfig":{"backendName":"mindieservice_llm_engine","modelInstanceNumber":1,"npuDeviceIds":[[0,1]],"tokenizerProcessNumber":8,"multiNodesInferEnabled":false,"multiNodesInferPort":1120,"interNodeTLSEnabled":true,"interNodeTlsCaPath":"security/grpc/ca/","interNodeTlsCaFiles":["ca.pem"],"interNodeTlsCert":"security/grpc/certs/server.pem","interNodeTlsPk":"security/grpc/keys/server.key.pem","interNodeTlsPkPwd":"security/grpc/pass/mindie_server_key_pwd.txt","interNodeTlsCrlPath":"security/grpc/certs/","interNodeTlsCrlFiles":["server_crl.pem"],"interNodeKmcKsfMaster":"tools/pmt/master/ksfa","interNodeKmcKsfStandby":"tools/pmt/standby/ksfb","ModelDeployConfig":{"maxSeqLen":2560,"maxInputTokenLen":2048,"truncation":false,"ModelConfig":[{"modelInstanceType":"Standard","modelName":"DeepSeek-R1-Distill-Qwen-32B","modelWeightPath":"/root/work/DeepSeek-R1-Distill-Qwen-32B","worldSize":2,"cpuMemSize":5,"npuMemSize":-1,"backendType":"atb","trustRemoteCode":false}]},"ScheduleConfig":{"templateType":"Standard","templateName":"Standard_LLM","cacheBlockSize":128,"maxPrefillBatchSize":50,"maxPrefillTokens":8192,"prefillTimeMsPerReq":150,"prefillPolicyType":0,"decodeTimeMsPerReq":50,"decodePolicyType":0,"maxBatchSize":200,"maxIterTimes":512,"maxPreemptCount":0,"supportSelectBatch":false,"maxQueueDelayMicroseconds":5000}}}四、跑服务化(有个加载模型的过程需要点时间)
cd /usr/local/Ascend/mindie/latest/mindie-service/bin&&./mindieservice_daemon
成功标志：
Daemon start success!
4.1 命令行推理方式
新开一个terminal(问问题，可以不进docker)
curl 192.168.2.71:1040/generate -d ‘{
“prompt”: “请输出100个生僻字?”,
“max_tokens”: 32,
“stream”: false,
“do_sample”:true,
“repetition_penalty”: 1.00,
“temperature”: 0.01,
“top_p”: 0.001,
“top_k”: 1,
“model”: “qwen”
}’
大概3秒回答问题
五、MindIE+webUI方式
关闭防火墙（在docker外执行）：
systemctl stop firewalld
安装webUI:
新开一个terminal,进docker：
docker exec -it sakway bash
cd /root/work/
pip install open-webui --index-url https://mirrors.huaweicloud.com/repository/pypi/simple/
这里open-webui的安装，大概需要十来分钟
安装成功后，到跑服务化的界面按Ctrl+C停止服务化进程（mindieservice_daemon）：
vim /usr/local/Ascend/mindie/latest/mindie-service/conf/config.json
将"ipAddress" : “127.0.0.1”,修改为实际IP地址"ipAddress" : “192.168.2.71”,
如果已经是"ipAddress" : "192.168.2.71"则不需要再修改
启动Open-WebUI服务：
open-webui serve
成功标志：
有一个大的OPEN WEBUI的LOGO
新开一个terminal,进docker：
docker exec -it sakway bash
跑服务化(有个加载模型的过程需要点时间)：
cd /usr/local/Ascend/mindie/latest/mindie-service/bin&&./mindieservice_daemon
成功标志：
Daemon start success!
在web浏览器中访问：
http://192.168.2.71:8080
点击开始使用
首次需要创建管理员账户：
名称：sakway
电子邮箱：[email protected]
密码：ABC123
点击创建管理员账户，此时会提示注册成功，已登录
点击确认，开始使用
点击右上角带颜色的圆圈图标（选择管理员面板）
点击上面那一排右边的设置
点击左侧的外部链接
将"管理OpenAI API连接"修改为实际的IP
https://api.openai.com/v1修改为http://192.168.2.71:1040/v1
点击该行最右边的"设置"，点击刷新，在弹出的“编辑连接”页面中点击保存
新开一个浏览器在web上访问http://192.168.2.71:8080开启对话
该局域网内的其他的PC用户也可以在浏览器打开http://192.168.2.71:8080开启新对话
以上大语言模型顺利跑成功
############################################################################################
