結(jié)論:

• Windows 無法通過WLS2的AMD ROCm Docker無法找到GPU, 參考: https://unix.stackexchange.com/questions/715847/wsl2-issue-installing-new-kernel/715922#715922
• Linux作為宿主機(jī)應(yīng)該是可以的, 參考: https://github.com/harakas/amd_igpu_yolo_v8
  這里記錄一些WLS下嘗試過程.

啟動(dòng)Docker deskop

啟動(dòng)Docker deskop, 通過界面可設(shè)置 http/https proxy, 以加快docker鏡像源的訪問.

準(zhǔn)備 Dockerfile

參考 https://github.com/harakas/amd_igpu_yolo_v8 , 該reps 使用的Linux作為宿主機(jī).
在 d:\my_workspace\docker\ 創(chuàng)建一個(gè) Dockfile 文件:

# Based on
#   https://rocm.docs.amd.com/projects/install-on-linux/en/latest/how-to/3rd-party/pytorch-install.html#using-docker-with-pytorch-pre-installed
#   https://pytorch.org/hub/ultralytics_yolov5/

FROM rocm/pytorch:rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1

RUN apt-get -y update
RUN apt-get -y upgrade

RUN apt-get install -y migraphx
RUN apt-get install -y fonts-freefont-ttf

RUN pip install -U pip
RUN pip install -U 'ultralytics' 'gitpython>=3.1.30' 'Pillow>=10.0.1' 'numpy>=1.23.5' 'scipy>=1.11.4' 'onnx>=1.12.0' 'onnxruntime' 

RUN mkdir /opt/cwd

WORKDIR /opt/cwd

ENTRYPOINT ["/opt/conda/envs/py_3.10/bin/python"]

基于該 Dockfile 文件創(chuàng)建鏡像

docer 鏡像 tag 為 rocm-pytorch , 我編譯耗時(shí)近兩個(gè)小時(shí).

cd D:\my_workspace\docker
d:
"C:\Program Files\Docker\Docker\resources\bin\docker.exe" build -t rocm-pytorch .

運(yùn)行docker鏡像

創(chuàng)建 bash 文件 docker_run.sh , 內(nèi)容如下, 然后運(yùn)行

#! /bin/bash
#
# Based on
#   https://rocm.docs.amd.com/projects/install-on-linux/en/latest/how-to/3rd-party/pytorch-install.html#using-docker-with-pytorch-pre-installed
#
# The following environment variables are necessary for ROCM to work properly:
#   -e HSA_ENABLE_SDMA=0
#       So that ROCM would work at all due to missing pce atomics
#   -e HSA_OVERRIDE_GFX_VERSION=9.0.0
#       ROCM/Tensile, etc do no work with gfx90c due to missing profile files
#       I have gfx90c so I override the gfx version for the libraries to work properly
#       See https://github.com/ROCm/ROCm/issues/1743#issuecomment-1149902796
#
# Also you need the following groups for ROCM to run properly (as we do not run as root):
#   --group-add video --group-add _ssh --group-add render
#

docker run -it \
  --cap-add=SYS_PTRACE \
  --security-opt seccomp=unconfined \
  --privileged \
  --device=/dev/kfd \
  --device=/dev/dri \
  --group-add video \
  --group-add _ssh \
  --group-add render \  
  -e HSA_ENABLE_SDMA=0 \
  -e HSA_OVERRIDE_GFX_VERSION=9.0.0 \
  -e PYTHONPATH=/opt/rocm-5.7.0/lib/ \
  --ipc=host \
  --shm-size 16G \
  --name my_docker \    
  -v /d/my_workspace/docker_vol:/docker_vol \  
  rocm-pytorch 

說明:

• 參數(shù)--shm-size是Docker鏡像的共享內(nèi)存大小（Shared Memories Size),如果需要訓(xùn)練很大的模型，可以根據(jù)實(shí)際需要調(diào)大。
• 將 d:\my_workspace\docker_vol 目錄映射到docker中
• 鏡像中已經(jīng)創(chuàng)建了 /opt/cwd 作為工作目錄
• 也可以增加 --network=host 參數(shù)試試
• 也可以增加 - e DEVICE=cuda 參數(shù)試試

當(dāng)創(chuàng)建 mydocker 容器后, 我們可以通過指定容器tag來運(yùn)行, 比如:

docker container start my_docker
docker container stop my_docker
docker container rm my_docker

python中是否能檢測(cè)出假冒的Cuda

命令行驗(yàn)證:

python3 -c 'import torch' 2> /dev/null && echo 'Success' || echo 'Failure'
python3 -c 'import torch; print(torch.cuda.is_available())'

python 解釋器中驗(yàn)證:

import torch
torch.cuda.is_available()

結(jié)果顯示False, 也就是沒有被認(rèn)為是CUDA, 在docker中執(zhí)行 rocminfo 命令, 顯示:

# 執(zhí)行 rocminfo
ROCk module is NOT loaded, possibly no GPU devices
# 執(zhí)行 sudo modprobe amdgpu
modprobe: FATAL: Module amdgpu not found in directory /lib/modules/5.15.133.1-microsoft-standard-WSL2

參考: https://unix.stackexchange.com/questions/715847/wsl2-issue-installing-new-kernel/715922#715922 , 發(fā)現(xiàn)該問題無法解決, 放棄繼續(xù)探索