作者： 张强

一 环境

1. 处理器： 鲲鹏920新型号处理器（7270Z、7280Z、7250Y、7260Y等)
2. Host OS: openEuler24.03 SP2
3. Guetst OS: openEuler24.03 SP2

二 模型

1. 测试数据集： ShareGPT_V3_unfiltered_cleaned_split.json
2. 大模型：Llama-2-7b-hf

三 Host 环境搭建

1. 安装依赖

yum install libvirt qemu edk2-aarch64 qemu-nbd -y

2. 将edk2中的QEMU_EFI.fd对齐到64M 

fallocate -l $((64 * 1024 * 1024)) /usr/share/edk2/aarch64/QEMU_EFI.fd

3. 修改内核启动参数

vim /boot/efi/EFI/openEuler/grub.cfg

在内核启动参数 linux /vmlinuz-xxx末尾添加 

virtcca_cvm_host=1 arm_smmu_v3.disable_ecmdq=1 vfio_pci.disable_idle_d3=1

四 机密虚机环境搭建

1 下载虚机镜像

wget https://repo.openeuler.org/openEuler-24.03-LTS-SP2/virtual_machine_img/aarch64/openEuler-24.03-LTS-SP2-aarch64.qcow2.xz
xz -d openEuler-24.03-LTS-SP2-aarch64.qcow2.xz

2. 修改内核参数

modprobe nbd max_part=16
qemu-nbd -c /dev/nbd0 openEuler-24.03-LTS-SP2-aarch64.qcow2
mkdir /mnt/qcow2
mount /dev/nbd0p1 /mnt/qcow2
vim /mnt/qcow2/efi/EFI/openEuler/grub.cfg
# 在启动参数末尾加上
cma=64M virtcca_cvm_guest=1 cvm_guest=1 swiotlb=65536,force loglevel=8
umount /mnt/qcow2
qemu-nbd -d /dev/nbd0

3. 增加虚拟网桥
新建network.xml, 写入如下内容

<network>
  <name>default</name>
  <uuid>c256bbb0-408d-47b4-83f7-b8e4085cf85c</uuid>
  <forward mode='nat'/>
  <bridge name='virbr0' stp='on' delay='0'/>
  <mac address='52:54:00:bb:eb:bf'/>
  <ip address='192.168.122.1' netmask='255.255.255.0'>
    <dhcp>
      <range start='192.168.122.2' end='192.168.122.254'/>
    </dhcp>
  </ip>
</network>

virsh net-define network.xml
virsh net-start default

4. 将GPU直通到机密虚机中
lspci |grep -i xxx查看BDF号，写入，libvirt配置文件中
5. 启动机密虚机
新建libvirt配置文件cvm-gpu.xml,写入如下内容

<domain type='kvm' xmlns:qemu='http://libvirt.org/schemas/domain/qemu/1.0'>
  <name>cvm-gpu</name>
  <memory unit='GiB'>16</memory>
  <currentMemory unit='GiB'>16</currentMemory>
  <vcpu placement='static'>16</vcpu>
  <iothreads>1</iothreads>

  <cputune>
    <vcpupin vcpu='0' cpuset='0'/>
    <vcpupin vcpu='1' cpuset='1'/>
    <vcpupin vcpu='2' cpuset='2'/>
    <vcpupin vcpu='3' cpuset='3'/>
    <vcpupin vcpu='4' cpuset='4'/>
    <vcpupin vcpu='5' cpuset='5'/>
    <vcpupin vcpu='6' cpuset='6'/>
    <vcpupin vcpu='7' cpuset='7'/>
    <vcpupin vcpu='8' cpuset='8'/>
    <vcpupin vcpu='9' cpuset='9'/>
    <vcpupin vcpu='10' cpuset='10'/>
    <vcpupin vcpu='11' cpuset='11'/>
    <vcpupin vcpu='12' cpuset='12'/>
    <vcpupin vcpu='13' cpuset='13'/>
    <vcpupin vcpu='14' cpuset='14'/>
    <vcpupin vcpu='15' cpuset='15'/>
    <emulatorpin cpuset='0-15'/>
  </cputune>
  <numatune>
    <memnode cellid='0' mode='strict' nodeset='0-3'/>
  </numatune>
  <resource>
    <partition>/machine</partition>
  </resource>
  <os>
    <type arch='aarch64' machine='virt-6.1'>hvm</type>
    <loader readonly='yes' type='rom'>/usr/share/edk2/aarch64/QEMU_EFI.fd</loader>
    <boot dev='hd'/>
  </os>
  <features>
    <gic version='3'/>
  </features>
  <cpu mode='host-passthrough' check='none'>
    <topology sockets='1' dies='1' clusters='1' cores='16' threads='1'/>
    <numa>
      <cell id='0' cpus='0-15' memory='16' unit='GiB'/>
    </numa>
  </cpu>
  <clock offset='utc'/>
  <on_poweroff>destroy</on_poweroff>
  <on_reboot>restart</on_reboot>
  <on_crash>destroy</on_crash>
  <devices>
          <emulator>/usr/libexec/qemu-kvm</emulator>
    <disk type='file' device='disk' model='virtio-non-transitional'>
      <driver name='qemu' type='qcow2' cache='none' queues='2' iommu='on'/>
      <source file='/home/test/qcow2/openEuler-24.03-LTS-SP1-aarch64-cvm.qcow2'/>
      <target dev='vda' bus='virtio'/>
      <address type='pci' domain='0x0000' bus='0x01' slot='0x00' function='0x0'/>
    </disk>
    <controller type='pci' index='0' model='pcie-root'/>
    <controller type='pci' index='1' model='pcie-root-port'>
      <model name='pcie-root-port'/>
      <target chassis='1' port='0x8'/>
      <address type='pci' domain='0x0000' bus='0x00' slot='0x01' function='0x0' multifunction='on'/>
    </controller>
    <controller type='pci' index='2' model='pcie-root-port'>
      <model name='pcie-root-port'/>
      <target chassis='2' port='0x9'/>
      <address type='pci' domain='0x0000' bus='0x00' slot='0x01' function='0x1'/>
    </controller>
    <controller type='pci' index='3' model='pcie-root-port'>
      <model name='pcie-root-port'/>
      <target chassis='3' port='0xa'/>
      <address type='pci' domain='0x0000' bus='0x00' slot='0x01' function='0x2'/>
    </controller>
    <controller type='pci' index='4' model='pcie-root-port'>
      <model name='pcie-root-port'/>
      <target chassis='4' port='0xb'/>
      <address type='pci' domain='0x0000' bus='0x00' slot='0x01' function='0x3'/>
    </controller>
    <serial type='pty'>
      <target type='system-serial' port='0'>
        <model name='pl011'/>
      </target>
    </serial>
    <console type='pty'>
      <target type='serial' port='0'/>
    </console>
    <audio id='1' type='none'/>
    <hostdev mode='subsystem' type='pci' managed='yes'>
      <driver name='vfio'/>
      <source>
        <!-- 这里写入 lspci |grep -i nvidia 查询到的BDF号-->
        <address domain='0x0000' bus='0xab' slot='0x00' function='0x0'/>
      </source>
      <address type='pci' domain='0x0000' bus='0x02' slot='0x00' function='0x0'/>
    </hostdev>
    <interface type='bridge'>
      <mac address='52:54:00:d0:a3:be'/>
      <source bridge='virbr0'/>
      <model type='virtio-non-transitional'/>
      <driver queues='4' iommu='on'/>
      <address type='pci' domain='0x0000' bus='0x03' slot='0x00' function='0x0'/>
    </interface>

  </devices>
  <launchSecurity type='cvm'/>
  <qemu:commandline>
    <qemu:arg value='-object'/>
    <qemu:arg value='tmm-guest,id=tmm0,sve-vector-length=128,num-pmu-counters=1'/>
  </qemu:commandline>
</domain>

virsh define cvm-gpu.xml # 定义机密虚机。
virsh start cvm-gpu --console # 输入账号密码后，进入机密虚机。

五 机密虚机内vLLM推理环境搭建

1. 驱动下载、安装

2. 计算平台下载、安装

3. conda 下载、安装

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-aarch64.sh

4. 安装python3.10

conda create -n py310 python=3.10
conda activate py310

5. vLLM 安装

wget https://download.pytorch.org/whl/nightly/cu126/torch-2.6.0.dev20250104%2Bcu126-cp310-cp310-linux_aarch64.whl
pip install torch-2.6.0.dev20250104
pip install compressed_tensors
git clone https://github.com/vllm-project/vllm.git
cd vllm
git checkout v0.6.6
python use_existing_torch.py
pip install -r requirements-build.txt
pip install -e . --no-build-isolation

六 模型推理测试

python3 benchmarks/benchmark_throughput.py --dataset /home/test/ShareGPT_V3_unfiltered_cleaned_split.json --model /home/test/Llama-2-7B-hf/ -tp 1