昇腾910B部署满血DeepSeek-R1(可推理版)

设备规划

4台910B物理服务器，我也不知道这个叫910B还是920B来着。就叫910B得了

[root@ds-3 ~]# cat /etc/os-release 
NAME="openEuler"
VERSION="22.03 LTS"
ID="openEuler"
VERSION_ID="22.03"
PRETTY_NAME="openEuler 22.03 LTS"
ANSI_COLOR="0;31"

[root@ds-3 ~]# uname -a
Linux ds-3 5.10.0-60.18.0.50.oe2203.aarch64 #1 SMP Wed Mar 30 02:43:08 UTC 2022 aarch64 aarch64 aarch64 GNU/Linux

[root@ds-3 ~]# lscpu
Architecture:           aarch64
  CPU op-mode(s):       64-bit
  Byte Order:           Little Endian
CPU(s):                 192
  On-line CPU(s) list:  0-191
Vendor ID:              HiSilicon
  BIOS Vendor ID:       HiSilicon
  Model name:           Kunpeng-920
    BIOS Model name:    HUAWEI Kunpeng 920 5250
    Model:              0
    Thread(s) per core: 1
    Core(s) per socket: 48
    Socket(s):          4
    Stepping:           0x1
    Frequency boost:    disabled
    CPU max MHz:        2600.0000
    CPU min MHz:        200.0000

ds-3是master节点

主机名称	ip地址规划	NPU卡IP地址
ds-3	10.82.27.3	10.82.29.17~24
ds-4	10.82.27.4	10.82.29.15~32
ds-5	10.82.27.5	10.82.29.33~40
ds-6	10.82.27.6	10.82.29.41~48

安装NPU驱动

所有节点操作：驱动在对应社区

# 安装驱动，部分软件会安装不上，可忽略
yum -y install dkms gcc linux-header kernel-dev kernel-headers
rpm -ivh Ascend-hdk-910b-npu-driver-24.1.0-1.aarch64.rpm

# 更新固件
./Ascend-hdk-910b-npu-firmware_7.5.0.3.220.run --check
# 可能提示需要重启
./Ascend-hdk-910b-npu-firmware_7.5.0.3.220.run --full
# 检查驱动情况
/usr/local/Ascend/driver/tools/upgrade-tool --device_index -1 --component -1 --version

设置NPU的IP地址和NPUping检测

ds-3: 10.82.29.17~24

ds-4: 10.82.29.25~32

ds-5: 10.82.29.33~40

ds-6: 10.82.29.41~48

每个节点需要操作

hccn_tool -i 0 -ip -s address 10.82.29.33 netmask 255.255.255.0
hccn_tool -i 1 -ip -s address 10.82.29.34 netmask 255.255.255.0
hccn_tool -i 2 -ip -s address 10.82.29.35 netmask 255.255.255.0
hccn_tool -i 3 -ip -s address 10.82.29.36 netmask 255.255.255.0
hccn_tool -i 4 -ip -s address 10.82.29.37 netmask 255.255.255.0
hccn_tool -i 5 -ip -s address 10.82.29.38 netmask 255.255.255.0
hccn_tool -i 6 -ip -s address 10.82.29.39 netmask 255.255.255.0
hccn_tool -i 7 -ip -s address 10.82.29.40 netmask 255.255.255.0

# 给卡配置IP地址
for i in {0..7}; do hccn_tool -i $i -gateway -s gateway 10.82.29.254; done
for i in {0..7}; do hccn_tool -i $i -netdetect -s address 10.82.29.254; done

# 检测ip
for i in {17..24}; do hccn_tool -i 0 -ping -g address 10.82.29.$i pkt 3; done

# 获取卡的ip
for i in {0..7};do hccn_tool -i $i -ip -g; done

# 检查NPU底层tls校验行为一致性，建议全0，如果未配置，会导致模型加载超时
for i in {0..7}; do hccn_tool -i $i -tls -g ; done | grep switch
# 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; done

rank_table_file.json

在宿主机所有节点使用此配置文件，注意文件权限

[root@ds-3 ~]# ll /data/rank_table_full.json
-rw-r----- 1 root root 3.3K Feb 20 21:31 /data/rank_table_full.json

{
    "version": "1.0",
    "server_count": "4",
    "server_list": [
        {
            "server_id": "10.82.27.3",
            "container_ip": "10.82.27.3",
            "device": [
                { "device_id": "0", "device_ip": "10.82.29.17", "rank_id": "0" },
                { "device_id": "1", "device_ip": "10.82.29.18", "rank_id": "1" },
                { "device_id": "2", "device_ip": "10.82.29.19", "rank_id": "2" },
                { "device_id": "3", "device_ip": "10.82.29.20", "rank_id": "3" },
                { "device_id": "4", "device_ip": "10.82.29.21", "rank_id": "4" },
                { "device_id": "5", "device_ip": "10.82.29.22", "rank_id": "5" },
                { "device_id": "6", "device_ip": "10.82.29.23", "rank_id": "6" },
                { "device_id": "7", "device_ip": "10.82.29.24", "rank_id": "7" }
            ]
        },
        {
            "server_id": "10.82.27.4",
            "container_ip": "10.82.27.4",
            "device": [
                { "device_id": "0", "device_ip": "10.82.29.25", "rank_id": "8" },
                { "device_id": "1", "device_ip": "10.82.29.26", "rank_id": "9" },
                { "device_id": "2", "device_ip": "10.82.29.27", "rank_id": "10" },
                { "device_id": "3", "device_ip": "10.82.29.28", "rank_id": "11" },
                { "device_id": "4", "device_ip": "10.82.29.29", "rank_id": "12" },
                { "device_id": "5", "device_ip": "10.82.29.30", "rank_id": "13" },
                { "device_id": "6", "device_ip": "10.82.29.31", "rank_id": "14" },
                { "device_id": "7", "device_ip": "10.82.29.32", "rank_id": "15" }
            ]
        },
        {
            "server_id": "10.82.27.5",
            "container_ip": "10.82.27.5",
            "device": [
                { "device_id": "0", "device_ip": "10.82.29.33", "rank_id": "16" },
                { "device_id": "1", "device_ip": "10.82.29.34", "rank_id": "17" },
                { "device_id": "2", "device_ip": "10.82.29.35", "rank_id": "18" },
                { "device_id": "3", "device_ip": "10.82.29.36", "rank_id": "19" },
                { "device_id": "4", "device_ip": "10.82.29.37", "rank_id": "20" },
                { "device_id": "5", "device_ip": "10.82.29.38", "rank_id": "21" },
                { "device_id": "6", "device_ip": "10.82.29.39", "rank_id": "22" },
                { "device_id": "7", "device_ip": "10.82.29.40", "rank_id": "23" }
            ]
        },
        {
            "server_id": "10.82.27.6",
            "container_ip": "10.82.27.6",
            "device": [
                { "device_id": "0", "device_ip": "10.82.29.41", "rank_id": "24" },
                { "device_id": "1", "device_ip": "10.82.29.42", "rank_id": "25" },
                { "device_id": "2", "device_ip": "10.82.29.43", "rank_id": "26" },
                { "device_id": "3", "device_ip": "10.82.29.44", "rank_id": "27" },
                { "device_id": "4", "device_ip": "10.82.29.45", "rank_id": "28" },
                { "device_id": "5", "device_ip": "10.82.29.46", "rank_id": "29" },
                { "device_id": "6", "device_ip": "10.82.29.47", "rank_id": "30" },
                { "device_id": "7", "device_ip": "10.82.29.48", "rank_id": "31" }
            ]
        }
    ],
    "status": "completed"
}

下载权重/模型

国外地址：https://huggingface.co/deepseek-ai/DeepSeek-R1/tree/main

国内地址：https://modelers.cn/spaces/State_Cloud/DeepSeek-R1/tree/main

NPU侧权重转换 目前npu转换脚本不会自动复制tokenizer等文件
git clone https://gitee.com/ascend/ModelZoo-PyTorch.git
cd ModelZoo-PyTorch\MindIE\LLM\DeepSeek\DeepSeek-V2\NPU_inference
python fp8_cast_bf16.py --input-fp8-hf-path /path/to/DeepSeek-R1 --output-bf16-hf-path /path/to/deepseek-R1-bf16

所有节点：在宿主机查看转换后的权重路径，文件大小为1.3T。注意设置权重权限，否则后期启动模型报错

docker容器NPU配置

运行容器

ds-3: 10.82.29.17~24

ds-4:10.82.29.25~32

ds-5: 10.82.29.33~40

ds-6: 10.82.29.41~48

官方的服务化测试和模型测试可以跳过，镜像在华为的社区，需要权限才能下载，大小为14G

#!/bin/bash
docker stop DeepSeek-R1-full
docker stop DeepSeek-R1-full
docker run -itd \
  --privileged \
  --name=DeepSeek-R1-full \
  --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 /etc/localtime:/etc/localtime \
  -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 /data:/data \
  -e ATB_LLM_HCCL_ENABLE=1 \
  -e ATB_LLM_COMM_BACKEND="hccl" \
  -e HCCL_CONNECT_TIMEOUT=7200 \
  -e WORLD_SIZE=32 \
  -e HCCL_EXEC_TIMEOUT=0 \
  -e PYTORCH_NPU_ALLOC_CONF=expandable_segments:True \
  -e RANKTABLEFILE=/data/rank_table.json \
  -e MIES_CONTAINER_IP=`hostname -I |awk '{print $1}'` \
  -e OMP_NUM_THREADS=1 \
  -e NPU_MEMORY_FRACTION=0.95 \
  swr.cn-south-1.myhuaweicloud.com/ascendhub/mindie:2.0.T3-800I-A2-py311-openeuler24.03-lts \
  bash
  
docker exec DeepSeek-R1 /bin/bash -c \
"mkdir -p /usr/local/Ascend/mindie/latest/mindie-service/conf; \
cat > /usr/local/Ascend/mindie/latest/mindie-service/conf/config.json <<'EOF'
{
    "Version" : "1.0.0",
    "LogConfig" :
    {
        "logLevel" : "Info",
        "logFileSize" : 20,
        "logFileNum" : 20,
        "logPath" : "logs/mindie-server.log"
    },

    "ServerConfig" :
    {
        "ipAddress" : "10.82.27.3",
        "managementIpAddress" : "10.82.27.3",
        "port" : 1025,
        "managementPort" : 1026,
        "metricsPort" : 1027,
        "allowAllZeroIpListening" : false,
        "maxLinkNum" : 300,
        "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" : false,
        "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" : [[0,1,2,3,4,5,6,7]],
        "tokenizerProcessNumber" : 8,
        "multiNodesInferEnabled" : true,
        "multiNodesInferPort" : 1120,
        "interNodeTLSEnabled" : false,
        "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" : 10000,
            "maxInputTokenLen" : 2048,
            "truncation" : true,
            "ModelConfig" : [
                {
                    "modelInstanceType" : "Standard",
                    "modelName" : "deepseekr1",
                    "modelWeightPath" : "/data/DeepSeek-R1-bf16",
                    "worldSize" : 8,
                    "cpuMemSize" : 5,
                    "npuMemSize" : -1,
                    "backendType" : "atb",
                    "trustRemoteCode" : false
                }
            ]
        },

        "ScheduleConfig" :
        {
            "templateType" : "Standard",
            "templateName" : "Standard_LLM",
            "cacheBlockSize" : 128,

            "maxPrefillBatchSize" : 8,
            "maxPrefillTokens" : 2048,
            "prefillTimeMsPerReq" : 150,
            "prefillPolicyType" : 0,

            "decodeTimeMsPerReq" : 50,
            "decodePolicyType" : 0,

            "maxBatchSize" : 8,
            "maxIterTimes" : 1024,
            "maxPreemptCount" : 0,
            "supportSelectBatch" : false,
            "maxQueueDelayMicroseconds" : 5000
        }
    }
}
EOF"

启动模型

所有节点操作：docker exec -it DeepSeek-R1 bash 。注意权重权限, rank_table_full.json

cd /usr/local/Ascend/mindie/latest/mindie-service/
./bin/mindieservice_daemon

其他节点的启动日志

显存占用明显升高

验证测试

[root@ds-3 ~]# curl 10.82.27.3:1025/generate -X POST -d '{"inputs":"题目：糖果的数量，小明、小红和小刚一共有24颗糖果。已知：小明比小红多2颗糖果；小红比小刚少4颗糖果。问题：小明、小红和小刚分别有多少颗糖果？","parameters":{"max_new_tokens":500},"temperature":0.3, "top_p":0.3, "top_k":5, "do_sample":true, "repetition_penalty":1.05, "seed":128}'
{"generated_text":"答案：小明有8颗，小红有6颗，小刚有10颗。解析：设小红有x颗糖果，则小明有x+2颗，小刚有x+4颗。根据总数为24颗，建立方程x + (x+2) + (x+4) = 24，解得x=6。因此，小明8颗，小红6颗，小刚10颗。<｜end▁of▁sentence｜>"}[root@ds-3 ~]# 

ST -d ‘{“inputs”:“题目：糖果的数量，小明、小红和小刚一共有24颗糖果。已知：小明比小红多2颗糖果；小红比小刚少4颗糖果。问题：小明、小红和小刚分别有多少颗糖果？”,“parameters”:{“max_new_tokens”:500},“temperature”:0.3, “top_p”:0.3, “top_k”:5, “do_sample”:true, “repetition_penalty”:1.05, “seed”:128}’
{“generated_text”:“答案：小明有8颗，小红有6颗，小刚有10颗。解析：设小红有x颗糖果，则小明有x+2颗，小刚有x+4颗。根据总数为24颗，建立方程x + (x+2) + (x+4) = 24，解得x=6。因此，小明8颗，小红6颗，小刚10颗。<｜end▁of▁sentence｜>”}[root@ds-3 ~]#
``