Atlas 800T A2
华为系列

Atlas 800T A2

配置

华为 Atlas 800I A2推理服务器完整硬件配置表

核心硬件配置

模块配置参数
处理器4*鲲鹏920处理器（7265/5250型号），双路 Hydra 总线互连（30GT/s），集成 PCIe 4.0控制器
内存（512GB）32个 DDR4插槽，最高支持3200 MT/s，单条容量16/32/64GB
NPU 配置- 型号：昇腾910B3 AI 处理器
- 数量：8个 NPU 模组全互联架构
- 运行内存：单卡64GB/整机512GB
- 单卡性能：FP16算力280 TFLOPS
- 互联带宽：整机392GB/s
存储配置- 选项1：82.5英寸 SATA（槽位0-7）+22.5英寸 NVMe（槽位8-9）
- 选项2：42.5英寸 SATA（槽位0-1、4-5）+62.5英寸 NVMe（槽位2-3、6-9）
（SP686C RAID 卡支持 SATA/NVMe 混合使用）
网络接口- 参数面：8200GE QSFP 接口（RoCE 协议）
- 存储面/业务面：各225GE
- 管理面：1*1GE
- PCIe 互联：NPU 与 CPU 通过 PCIe 4.0 x16互联（总带宽512GB/s）

物理与扩展特性

模块配置参数
PCIe 扩展- 存储配置选项1支持3个 PCIe 4.0插槽
- 选项2支持2个 PCIe 4.0插槽
电源与散热4个热插拔2.6kW 电源（2+2冗余）
8个热插拔风扇（N+1冗余）
物理规格4U 机架式（175mm×447mm×790mm），净重70kg
工作环境温度5℃～35℃

关键功能说明

  1. NPU 逻辑映射：iBMC 显示 NPU18对应 npu-smi 工具的 NPU07，需注意物理位置与逻辑 ID 匹配
  2. 容器支持：通过 npu-smi 工具动态分配 AI CPU 和算力资源，支持物理机/虚拟机/容器化部署
  3. RAID 兼容性：仅 SP686C RAID 卡支持 SATA/NVMe 混合使用，其他 RAID 卡仅支持 NVMe 硬盘
  4. 视频处理能力：支持1080P 视频解码480FPS，JPEG 解码最大分辨率16384x16384

注意：昇腾910b3不支持容器共享技术

mindie 部署

docker-compose. yml

services:
  mindie:
    container_name: mindie
    image: llama3.1:8bv1  # 原始镜像:swr.cn-central-221.ovaijisuan.com/wh-aicc-fae/mindie:910B-ascend_24.1.rc3-cann_8.0.rc3.beta1-py_3.10-ubuntu_20.04-aarch64-mindie_1.0.RC3.04
    privileged: true
    networks:  
      - docker_default  
    volumes:
      - /usr/local/Ascend/driver:/usr/local/Ascend/driver:ro
      - /usr/local/Ascend/add-ons/:/usr/local/Ascend/add-ons/:ro 
      - /usr/local/sbin/npu-smi:/usr/local/sbin/npu-smi:ro
      - /usr/local/sbin/:/usr/local/sbin/:ro
      - /var/log/npu/conf/slog/slog.conf:/var/log/npu/conf/slog/slog.conf:ro
      - /var/log/npu/slog/:/var/log/npu/slog:ro
      - /var/log/npu/profiling/:/var/log/npu/profiling:ro 
      - /var/log/npu/dump/:/var/log/npu/dump:ro
      - /var/log/npu/:/usr/slog:ro
      - /share-storage/llmfinetuning/ms_cache:/root/.cache/modelscope
      - /share-storage/mindie/home/Meta-Llama-3.1-8B-Instruct:/home/model/Meta-Llama-3.1-8B-Instruct
      - /share-storage/mindie/home/law_sft:/home/model/law_sft
      - /share-storage/mindie/home/config.json:/usr/local/Ascend/mindie/latest/mindie-service/conf/config.json
      # - /share-storage/mindie/home/start.sh:/home/start.sh  # 添加 start.sh 文件映射
    expose:
      - 1025
    ports:
      - 1025:1025
    ipc: host
    tty: true
    shm_size: '16gb'
    # command: bash /home/start.sh  # 修改启动脚本为 start.sh
    devices:
      - /dev/davinci1
      - /dev/davinci2
      - /dev/davinci_manager
      - /dev/devmm_svm
      - /dev/hisi_hdc
    restart: unless-stopped

networks:  
  docker_default:  
    external: true

config. json

{
    "Version" : "1.0.0",
    "LogConfig" :
    {
        "logLevel" : "Info",
        "logFileSize" : 20,
        "logFileNum" : 20,
        "logPath" : "logs/mindservice.log"
    },

    "ServerConfig" :
    {
        "ipAddress" : "172.26.0.10",  // 只能填写ip地址，0.0.0.0无效
        "managementIpAddress" : "127.0.0.2",
        "port" : 1025,
        "managementPort" : 1026,
        "metricsPort" : 1027,
        "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",
        "tlsCrl" : "security/certs/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",
        "managementTlsCrl" : "security/certs/management/server_crl.pem",
        "kmcKsfMaster" : "tools/pmt/master/ksfa",
        "kmcKsfStandby" : "tools/pmt/standby/ksfb",
        "inferMode" : "standard",
        "interCommTLSEnabled" : false,
        "interCommPort" : 1121,
        "interCommTlsCaFile" : "security/grpc/ca/ca.pem",
        "interCommTlsCert" : "security/grpc/certs/server.pem",
        "interCommPk" : "security/grpc/keys/server.key.pem",
        "interCommPkPwd" : "security/grpc/pass/key_pwd.txt",
        "interCommTlsCrl" : "security/certs/server_crl.pem",
        "openAiSupport" : "vllm"
    },

    "BackendConfig" : {
        "backendName" : "mindieservice_llm_engine",
        "modelInstanceNumber" : 1,  // 服务实例个数，对应下一行第一层的个数
        "npuDeviceIds" : [[1,2]],  // 第一层表示实例，第二层表示每个实例GPU的个数
        "tokenizerProcessNumber" : 8,
        "multiNodesInferEnabled" : false,
        "multiNodesInferPort" : 1120,
        "interNodeTLSEnabled" : true,
        "interNodeTlsCaFile" : "security/grpc/ca/ca.pem",
        "interNodeTlsCert" : "security/grpc/certs/server.pem",
        "interNodeTlsPk" : "security/grpc/keys/server.key.pem",
        "interNodeTlsPkPwd" : "security/grpc/pass/mindie_server_key_pwd.txt",
        "interNodeTlsCrl" : "security/grpc/certs/server_crl.pem",
        "interNodeKmcKsfMaster" : "tools/pmt/master/ksfa",
        "interNodeKmcKsfStandby" : "tools/pmt/standby/ksfb",
        "ModelDeployConfig" :
        {
            "maxSeqLen" : 131072,  // 最大序列长度
            "maxInputTokenLen" : 131072,  // 输入的最大序列长度
            "truncation" : false,  // true表示超过输入最大序列长度，则按照一定的规则进行截取。
            "ModelConfig" : [
                {
                    "modelInstanceType" : "Standard",
                    "modelName" : "llama3.1_8b",
                    "modelWeightPath" : "/home/model/law_sft/",
                    "worldSize" : 2,  // 一个实例所拥有的NPU数量
                    "cpuMemSize" : 64,
                    "npuMemSize" : -1,
                    "backendType" : "atb"
                }
            ]
        },

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

            "maxPrefillBatchSize" : 50,
            "maxPrefillTokens" : 131072,  // 如果显存不够，降低这个参数可以降低显存。
            "prefillTimeMsPerReq" : 150,
            "prefillPolicyType" : 0,

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

            "maxBatchSize" : 200,
            "maxIterTimes" : 512,
            "maxPreemptCount" : 0,
            "supportSelectBatch" : false,
            "maxQueueDelayMicroseconds" : 5000
        }
    }
}

启动

source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/mindie/set_env.sh

nohup /usr/local/Ascend/mindie/latest/mindie-service/bin/mindieservice_daemon &

text-embeding-inference 部署

docker-compose. yml

services:
  bge-m3:
    container_name: bge-m3
    image: swr.cn-south-1.myhuaweicloud.com/ascendhub/mis-tei:6.0.RC3-800I-A2-aarch64
    entrypoint: /home/HwHiAiUser/start.sh
    command: 
      - "BAAI/bge-m3"
      - "0.0.0.0"
      - "8080"
    networks:
      - docker_default
    volumes:
      - /usr/local/Ascend/driver:/usr/local/Ascend/driver:ro
      - /usr/local/Ascend/add-ons/:/usr/local/Ascend/add-ons/:ro 
      - /usr/local/sbin/npu-smi:/usr/local/sbin/npu-smi:ro
      - /usr/local/sbin/:/usr/local/sbin/:ro
      - /var/log/npu/conf/slog/slog.conf:/var/log/npu/conf/slog/slog.conf:ro
      - /var/log/npu/slog/:/var/log/npu/slog:ro
      - /var/log/npu/profiling/:/var/log/npu/profiling:ro 
      - /var/log/npu/dump/:/var/log/npu/dump:ro
      - /var/log/npu/:/usr/slog:ro
      - /share-storage/mindie/model:/home/HwHiAiUser/model
    expose:
      - 8080
      - 9090
    ipc: host
    tty: true
    shm_size: '16gb'
    # stdin_open: true
    # command: lmdeploy serve api_server --backend pytorch --device ascend --eager-mode --chat-template llama3_1 ./sftmodels/identity
    environment:
      - ASCEND_VISIBLE_DEVICES=1 
    devices:
      - /dev/davinci0
      # - /dev/davinci1
      # - /dev/davinci2
      # - /dev/davinci3
      # - /dev/davinci4
      # - /dev/davinci5
      # - /dev/davinci6
      # - /dev/davinci7
      - /dev/davinci_manager
      - /dev/devmm_svm
      - /dev/hisi_hdc
    restart: unless-stopped


networks:
  docker_default:
    external: true

启动指令

/home/HwHiAiUser/start.sh BAAI/bge-m3 0.0.0.0 8080
/home/HwHiAiUser/start.sh BAAI/bge-reranker-v2-m3 0.0.0.0 9090

lmdeploy+nginx 负载均衡

docker-compose. yml

services:
  lmdeploy-1:
    container_name: lmdeploy-1
    image: lmdeploy-aarch64-ascend:latest
    networks:
      - docker_default
    volumes:
      - /share-storage/llmfinetuning/output:/opt/lmdeploy/sftmodels:ro
      - /usr/local/Ascend/driver:/usr/local/Ascend/driver:ro
      - /usr/local/Ascend/add-ons/:/usr/local/Ascend/add-ons/:ro
      - /usr/local/sbin/npu-smi:/usr/local/sbin/npu-smi:ro
      - /usr/local/sbin/:/usr/local/sbin/:ro
      - /var/log/npu/conf/slog/slog.conf:/var/log/npu/conf/slog/slog.conf:ro
      - /var/log/npu/slog/:/var/log/npu/slog:ro
      - /var/log/npu/profiling/:/var/log/npu/profiling:ro
      - /var/log/npu/dump/:/var/log/npu/dump:ro
      - /var/log/npu/:/usr/slog:ro
    expose:
      - 23333
    ipc: host
    tty: true
    shm_size: '16gb'
    # stdin_open: true
    # working_dir: /opt/lmdeploy
    # environment:
    #   - ASCEND_VISIBLE_DEVICES=2
    # entrypoint: lmdeploy
    # command:
    #   - "serve"
    #   - "api_server"
    #   - "--backend"
    #   - "pytorch"
    #   - "--device"
    #   - "ascend"
    #   - "--eager-mode"
    #   - "--model-name"
    #   - "llama3_1_8b"
    #   - "--chat-template"
    #   - "llama3_1"
    #   - "./sftmodels/Meta-Llama-3.1-8B-Instruct"
    # command: lmdeploy serve api_server --backend pytorch --device ascend --eager-mode --chat-template llama3_1 ./sftmodels/identity
    devices:
      # - /dev/davinci0
      # - /dev/davinci1
      - /dev/davinci2
      # - /dev/davinci3
      # - /dev/davinci4
      # - /dev/davinci5
      # - /dev/davinci6
      # - /dev/davinci7
      - /dev/davinci_manager
      - /dev/devmm_svm
      - /dev/hisi_hdc
    restart: unless-stopped

  lmdeploy-2:
    container_name: lmdeploy-2
    image: lmdeploy-aarch64-ascend:latest
    networks:
      - docker_default
    volumes:
      - /share-storage/llmfinetuning/output:/opt/lmdeploy/sftmodels:ro
      - /usr/local/Ascend/driver:/usr/local/Ascend/driver:ro
      - /usr/local/Ascend/add-ons/:/usr/local/Ascend/add-ons/:ro
      - /usr/local/sbin/npu-smi:/usr/local/sbin/npu-smi:ro
      - /usr/local/sbin/:/usr/local/sbin/:ro
      - /var/log/npu/conf/slog/slog.conf:/var/log/npu/conf/slog/slog.conf:ro
      - /var/log/npu/slog/:/var/log/npu/slog:ro
      - /var/log/npu/profiling/:/var/log/npu/profiling:ro
      - /var/log/npu/dump/:/var/log/npu/dump:ro
      - /var/log/npu/:/usr/slog:ro
    expose:
      - 23333
    ipc: host
    tty: true
    shm_size: '16gb'
    # stdin_open: true
    working_dir: /opt/lmdeploy
    # environment:
    #   - ASCEND_VISIBLE_DEVICES=3
    # entrypoint: lmdeploy
    # command:
    #   - "serve"
    #   - "api_server"
    #   - "--backend"
    #   - "pytorch"
    #   - "--device"
    #   - "ascend"
    #   - "--eager-mode"
    #   - "--model-name"
    #   - "llama3_1_8b"
    #   - "--chat-template"
    #   - "llama3_1"
    #   - "./sftmodels/Meta-Llama-3.1-8B-Instruct"
    # command: lmdeploy serve api_server --backend pytorch --device ascend --eager-mode --chat-template llama3_1 ./sftmodels/identity
    devices:
      # - /dev/davinci0
      # - /dev/davinci1
      # - /dev/davinci2
      - /dev/davinci3
#       - /dev/davinci4
#       - /dev/davinci5
#       - /dev/davinci6
#       - /dev/davinci7
      - /dev/davinci_manager
      - /dev/devmm_svm
      - /dev/hisi_hdc
    restart: unless-stopped

  ngixn:
    image: nginx:latest
    container_name: lmdeploy-nginx
    networks:
      - docker_default
    volumes:
      - /share-storage/lmdeploy_nginx/nginx.conf:/etc/nginx/nginx.conf:ro
      - /share-storage/lmdeploy_nginx/datasets:/datasets:ro
    ports:
      - "23333:23333"
      - "23334:23334"


networks:
  docker_default:
    external: true

nginx. conf

user  nginx;

events {
    worker_connections   1000;
}

http {
    # 显示目录
    # autoindex on;
    # 显示文件大小
    # autoindex_exact_size on;
    # 显示文件时间
    # autoindex_localtime on;
    upstream lmdeploy {
        #ip_hash;
        server lmdeploy-1:23333;
        server lmdeploy-2:23333;
        # server lmdeploy-3:23333;
    }
    server {
        listen 23333;
        location / {
            proxy_buffering off;
            proxy_pass http://lmdeploy/;
        }
    }
    server {
        listen 23334;

        # 避免中文乱码
        charset utf-8;

        location / {
            root        /datasets;
            sendfile    on;
            autoindex   on;
            autoindex_exact_size  off;
            autoindex_localtime   on;
        }

    }
}

启动

lmdeploy serve api_server --backend pytorch --device ascend --eager-mode --chat-template llama3_1 ./sftmodels/identity
LICENSED UNDER CC BY-NC-SA 4.0