[AMD] podman 테스트

- OS: RHEL 9.4

- GPU: Radeon PRO W6800

 

1. Redhat OS local repo 생성

[root@cray ~]# cat /etc/yum.repos.d/local.repo  [media-baseos] name=BaseOS baseurl=file:///data/REPO/rhel9.4/BaseOS gpgcheck=1 enabled=1 gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-redhat-release [media-appstream] name=AppStream baseurl=file:///data/REPO/rhel9.4/AppStream gpgcheck=1 enabled=1 gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-redhat-release

 

2. AMD GPU driver & rocm download

- 아래 예제와 같이 사용하면 AMD에서 제공하는 repository에서 rpm파일들을 일괄 다운 받을 수 있다.

# wget -e robots=off -r -np 'https://repo.radeon.com/amdgpu/latest/rhel/9.4/main/x86_64' # wget -e robots=off -r -np 'https://repo.radeon.com/rocm/rhel9/latest/main'

- "-e robots=off" 옵션을 이용하면 크롤링 로봇을 차단하는 기능을 끌 수 있다. 

 

3. AMD GPU repo 생성

# yum install createrepo_c.x86_64 # cd /data/REPO/AMD/amdgpu # createrepo . # cd /data/REPO/AMD/rocm # createrepo . # vi /etc/yum.repos.d/amd.repo - - 아래 amd.repo 예제 참고 - -

 

- amd.repo 예제

[amdgpu] name=amdgpu Packages gpgcheck=0 enabled=1 baseurl=file:///data/REPO/AMD/amdgpu [rocm] name=rocm Packages gpgcheck=0 enabled=1 baseurl=file:///data/REPO/AMD/rocm

 

4. 패키지 설치

# yum groupinstall "Development Tools" # yum install amdgpu-dkms # yum install rocm

 

- 설치 후 lsmod 명령을 이용해 amdgpu 모듈이 잘 로드되어져 있는지 확인한다.

[sylee@cray ~]$ lsmod | grep amdgpu amdgpu              15462400  0 amddrm_ttm_helper      16384  1 amdgpu amdttm                106496  2 amdgpu,amddrm_ttm_helper amddrm_buddy           24576  1 amdgpu amdxcp                 16384  1 amdgpu i2c_algo_bit           16384  1 amdgpu drm_exec               16384  1 amdgpu drm_suballoc_helper    16384  1 amdgpu amd_sched              69632  1 amdgpu amdkcl                 32768  3 amd_sched,amdttm,amdgpu drm_display_helper    212992  1 amdgpu drm_kms_helper        245760  4 drm_display_helper,amdgpu video                  73728  1 amdgpu drm                   741376  11 drm_kms_helper,drm_exec,amd_sched,amdttm,drm_suballoc_helper,drm_display_helper,amdgpu,amddrm_buddy,amddrm_ttm_helper,amdxcp

 

5. GPU 권한 udev mode 0666

# vi /etc/udev/rules.d/70-amdgpu.rules KERNEL=="kfd", GROUP="video", MODE="0666" # udevadm control --reload-rules && udevadm trigger

- 참고: https://rocm.docs.amd.com/projects/install-on-linux/en/latest/install/prerequisites.html

 

6. selinux disabled

# grubby --update-kernel ALL --args selinux=0 # shutdown -r now

 

7. AMD GPU podman test

# yum install podman-docker.noarch podman-plugins.x86_64
# setsebool -P container_use_devices 1 

 

- container 에서 pytorch 실행 예제

$ podman pull docker.io/rocm/pytorch:latest $ podman run -it --device /dev/kfd --device /dev/dri --net=host --security-opt=no-new-privileges --cap-drop=ALL docker.io/rocm/pytorch:latest python3 >>> import torch; >>> torch.cuda.is_available(); True >>> torch.cuda.current_device(); 0 >>> torch.cuda.get_device_name(0); 'AMD Radeon Graphics'

 

- container 에서 rocm-smi 실행 예제

[sylee@cray ~]$ podman run -it --device /dev/kfd --device /dev/dri --net=host --security-opt=no-new-privileges --cap-drop=ALL docker.io/rocm/pytorch:latest rocm-smi ======================================== ROCm System Management Interface ======================================== ================================================== Concise Info ================================================== Device  Node  IDs              Temp    Power  Partitions          SCLK  MCLK   Fan    Perf  PwrCap  VRAM%  GPU%                 (DID,     GUID)  (Edge)  (Avg)  (Mem, Compute, ID)                                                  ================================================================================================================== 0       1     0x73a3,   45123  24.0°C  8.0W   N/A, N/A, 0         0Mhz  96Mhz  20.0%  auto  213.0W  0%     0%     ================================================================================================================== ============================================== End of ROCm SMI Log =============================================== [sylee@cray ~]$

- 참고: https://access.redhat.com/solutions/7073764