准备工作
- 显卡为8卡910B，aarch64架构鲲鹏cpu
[root@localhost ~]# lscpu
架构： aarch64
CPU 运行模式： 64-bit
字节序： Little Endian
CPU: 128
在线 CPU 列表： 0-127
厂商 ID： HiSilicon
BIOS Vendor ID: HiSilicon
型号名称： Kunpeng-920
BIOS Model name: HUAWEI Kunpeng 920 7260
型号： 0
每个核的线程数： 1
每个座的核数： 64
座： 2
步进： 0x1
BogoMIPS： 200.00
标记： fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma dcpop asimddp asimdfh
m ssbs
Caches (sum of all):
L1d: 8 MiB (128 instances)
L1i: 8 MiB (128 instances)
L2: 64 MiB (128 instances)
L3: 128 MiB (4 instances)
NUMA:
NUMA 节点： 4
NUMA 节点0 CPU： 0-31
NUMA 节点1 CPU： 32-63
NUMA 节点2 CPU： 64-95
NUMA 节点3 CPU： 96-127
Vulnerabilities:
Gather data sampling: Not affected
Itlb multihit: Not affected
L1tf: Not affected
Mds: Not affected
Meltdown: Not affected
Mmio stale data: Not affected
Retbleed: Not affected
Spec rstack overflow: Not affected
Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Spectre v1: Mitigation; __user pointer sanitization
Spectre v2: Not affected
Srbds: Not affected
Tsx async abort: Not affected[root@localhost ~]# npu-smi info
+------------------------------------------------------------------------------------------------+
| npu-smi 25.5.1 Version: 25.5.1 |
+---------------------------+---------------+----------------------------------------------------+
| NPU Name | Health | Power(W) Temp(C) Hugepages-Usage(page)|
| Chip | Bus-Id | AICore(%) Memory-Usage(MB) HBM-Usage(MB) |
+===========================+===============+====================================================+
| 3 910B4-1 | OK | 76.5 41 0 / 0 |
| 0 | 0000:2F:00.0 | 0 0 / 0 3202 / 65536 |
+===========================+===============+====================================================+
| 4 910B4-1 | OK | 77.4 41 0 / 0 |
| 0 | 0000:2C:00.0 | 0 0 / 0 3202 / 65536 |
+===========================+===============+====================================================+
| 5 910B4-1 | OK | 84.4 43 0 / 0 |
| 0 | 0000:2B:00.0 | 0 0 / 0 3202 / 65536 |
+===========================+===============+====================================================+
| 6 910B4-1 | OK | 79.7 42 0 / 0 |
| 0 | 0000:33:00.0 | 0 0 / 0 3203 / 65536 |
+===========================+===============+====================================================+
| 7 910B4-1 | OK | 87.0 44 0 / 0 |
| 0 | 0000:11:00.0 | 0 0 / 0 3203 / 65536 |
+===========================+===============+====================================================+
| 8 910B4-1 | OK | 77.7 42 0 / 0 |
| 0 | 0000:12:00.0 | 0 0 / 0 3202 / 65536 |
+===========================+===============+====================================================+
| 9 910B4-1 | OK | 76.0 41 0 / 0 |
| 0 | 0000:1A:00.0 | 0 0 / 0 3202 / 65536 |
+===========================+===============+====================================================+
| 10 910B4-1 | OK | 79.3 42 0 / 0 |
| 0 | 0000:19:00.0 | 0 0 / 0 3202 / 65536 |
+===========================+===============+====================================================+- 系统为openEuler 22.03 ARM架构
cat /etc/os-release
NAME="openEuler"
VERSION="22.03 (LTS-SP4)"
ID="openEuler"
VERSION_ID="22.03"
PRETTY_NAME="openEuler 22.03 (LTS-SP4)"
ANSI_COLOR="0;31"一、安装docker compose
修改安装源
cat > /etc/yum.repos.d/docker-ce.repo << 'EOF'
[docker-ce-stable]
name=Docker CE Stable - $basearch
baseurl=https://repo.huaweicloud.com/docker-ce/linux/centos/7/$basearch/stable
enabled=1
gpgcheck=1
gpgkey=https://repo.huaweicloud.com/docker-ce/linux/centos/gpg
EOF安装docker compose
dnf install -y docker-ce docker-ce-cli containerd.io docker-compose启动docker并开启开机自启动
systemctl start docker && systemctl enable docker二、下载安装910B固件和驱动程序
1、下载固件和驱动
下载页面社区版-固件与驱动-昇腾社区
过滤选项为 npu`run后缀`AArch64
下载链接
Atlas A2 中心推理和训练硬件产品ARM架构驱动包（run格式）
Atlas A2 中心推理和训练硬件产品固件包（run格式）
或用以下指令直接下载(可能会失效)：
wget https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/Ascend%20HDK/Ascend%20HDK%2025.5.1/Ascend-hdk-910b-npu-driver_25.5.1_linux-aarch64.run?response-content-type=application/octet-streamwget https://ascend-repo.obs.cn-east-2.myhuaweicloud.com/Ascend%20HDK/Ascend%20HDK%2025.5.1/Ascend-hdk-910b-npu-firmware_7.8.0.6.201.run?response-content-type=application/octet-stream2、安装依赖
yum install -y make dkms gcc kernel-headers-$(uname -r) kernel-devel-$(uname -r)3、创建华为用户组
groupadd HwHiAiUser && useradd -g HwHiAiUser -d /home/HwHiAiUser -m HwHiAiUser -s /bin/bash4、安装驱动
chmod +x * && ./Ascend-hdk-910b-npu-driver_25.5.1_linux-aarch64.run --full --install-for-all出现 Driver package installed successfully!为安装成功
5、安装固件
./Ascend-hdk-910b-npu-firmware_7.8.0.6.201.run --full出现如下所示为安装成功
[Firmware] [2026-07-02 10:57:41] [INFO]The firmware of [8] chips are successfully upgraded.
[Firmware] [2026-07-02 10:57:42] [INFO]Firmware package installed successfully! Reboot now or after driver installation for the installation/upgrade to take effect.安装完成后按照提示重启服务器
6、查看npu信息
npu-smi info如下所示表示显卡集群正常
[root@localhost ~]# npu-smi info
+------------------------------------------------------------------------------------------------+
| npu-smi 25.5.1 Version: 25.5.1 |
+---------------------------+---------------+----------------------------------------------------+
| NPU Name | Health | Power(W) Temp(C) Hugepages-Usage(page)|
| Chip | Bus-Id | AICore(%) Memory-Usage(MB) HBM-Usage(MB) |
+===========================+===============+====================================================+
| 3 910B4-1 | OK | 75.9 40 0 / 0 |
| 0 | 0000:2F:00.0 | 0 0 / 0 3202 / 65536 |
+===========================+===============+====================================================+
| 4 910B4-1 | OK | 76.9 41 0 / 0 |
| 0 | 0000:2C:00.0 | 0 0 / 0 3202 / 65536 |
+===========================+===============+====================================================+
| 5 910B4-1 | OK | 83.7 41 0 / 0 |
| 0 | 0000:2B:00.0 | 0 0 / 0 3203 / 65536 |
+===========================+===============+====================================================+
| 6 910B4-1 | OK | 79.1 40 0 / 0 |
| 0 | 0000:33:00.0 | 0 0 / 0 3202 / 65536 |
+===========================+===============+====================================================+
| 7 910B4-1 | OK | 86.0 43 0 / 0 |
| 0 | 0000:11:00.0 | 0 0 / 0 3202 / 65536 |
+===========================+===============+====================================================+
| 8 910B4-1 | OK | 77.2 40 0 / 0 |
| 0 | 0000:12:00.0 | 0 0 / 0 3203 / 65536 |
+===========================+===============+====================================================+
| 9 910B4-1 | OK | 75.5 40 0 / 0 |
| 0 | 0000:1A:00.0 | 0 0 / 0 3202 / 65536 |
+===========================+===============+====================================================+
| 10 910B4-1 | OK | 78.4 40 0 / 0 |
| 0 | 0000:19:00.0 | 0 0 / 0 3202 / 65536 |
+===========================+===============+====================================================+
+---------------------------+---------------+----------------------------------------------------+
| NPU Chip | Process id | Process name | Process memory(MB) |
+===========================+===============+====================================================+
| No running processes found in NPU 3 |
+===========================+===============+====================================================+
| No running processes found in NPU 4 |
+===========================+===============+====================================================+
| No running processes found in NPU 5 |
+===========================+===============+====================================================+
| No running processes found in NPU 6 |
+===========================+===============+====================================================+
| No running processes found in NPU 7 |
+===========================+===============+====================================================+
| No running processes found in NPU 8 |
+===========================+===============+====================================================+
| No running processes found in NPU 9 |
+===========================+===============+====================================================+
| No running processes found in NPU 10 |
+===========================+===============+====================================================+三、安装vLLM
1、创建配置文件
创建目录
mkdir -p /home/qwen3 && cd /home/qwen3编写容器内启动脚本 start.sh
touch start.sh && chmod +x start.sh && cat > start.sh << 'EOF'
export VLLM_USE_MODELSCOPE=True
export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export HCCL_BUFFSIZE=512
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=1
export TASK_QUEUE_ENABLE=1
vllm serve /data/model/Qwen3-VL-235B-A22B-Instruct-w8a8-QuaRot \
--host 0.0.0.0 \
--port 8000 \
--tensor-parallel-size 8 \
--data-parallel-size 1 \
--seed 1024 \
--served-model-name qwen3 \
--max-num-seqs 32 \
--max-model-len 131072 \
--max-num-batched-tokens 8096 \
--enable-expert-parallel \
--trust-remote-code \
--gpu-memory-utilization 0.95 \
--hf-overrides '{"rope_parameters": {"rope_type":"yarn","rope_theta":1000000,"factor":4,"original_max_position_embeddings":32768}}' \
--compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}' \
--async-scheduling \
--served-model-name qwen3 \
--enable-auto-tool-choice \
--tool-call-parser hermes
EOF编写docker-compose文件，以下配置的宿主机目录映射 /mnt/sdb:/data需调整为实际模型存放路径
touch output.log && cat > docker-compose.yml << 'EOF'
services:
qwen:
image: quay.io/ascend/vllm-ascend:v0.19.1rc1-openeuler
container_name: qwen
network_mode: host
privileged: true
shm_size: 500g
devices:
- /dev/davinci0:/dev/davinci0
- /dev/davinci1:/dev/davinci1
- /dev/davinci2:/dev/davinci2
- /dev/davinci3:/dev/davinci3
- /dev/davinci4:/dev/davinci4
- /dev/davinci5:/dev/davinci5
- /dev/davinci6:/dev/davinci6
- /dev/davinci7:/dev/davinci7
- /dev/davinci_manager:/dev/davinci_manager
- /dev/devmm_svm:/dev/devmm_svm
- /dev/hisi_hdc:/dev/hisi_hdc
volumes:
- /usr/local/dcmi:/usr/local/dcmi
- /usr/local/Ascend/driver/tools/hccn_tool:/usr/local/Ascend/driver/tools/hccn_tool
- /usr/local/bin/npu-smi:/usr/local/bin/npu-smi
- /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/
- /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info
- /etc/ascend_install.info:/etc/ascend_install.info
- /mnt/sdb:/data
- ./start.sh:/workspace/start.sh
command: /bin/sh -c "/workspace/start.sh"
EOF2、启动vLLM容器
使用以下指令启动vLLM
docker compose up -d使用logs指令查看容器日志
docker compose logs -f qwen首次启动大约10min，出现以下日志内容，即启动完成
(APIServer pid=225) INFO: Started server process [225]
(APIServer pid=225) INFO: Waiting for application startup.
(APIServer pid=225) INFO: Application startup complete.3、测试api
使用curl测试模型输出
curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer sk-LPyEjJE10FbwbEVyviyHMQ" -d '{"model": "qwen3", "messages": [{"role": "user", "content": "你好，请介绍一下你自己"}]}'