[ROCM] Unable to open /dev/kfd read-write: Permission denied

일반 사용자 계정으로 rocminfo 명령 실행 시 오류 해결 방법

 

오류 내용

[sylee@cray ~]$ rocminfo ROCk module is loaded Unable to open /dev/kfd read-write: Permission denied sylee is not member of "video" group, the default DRM access group. Users must be a member of the "video" group or another DRM access  group in order for ROCm applications to run successfully.

- 일반계정(예: sylee)으로 AMD GPU의 rocminfo 명령 실행 시 /dev/kfd 디바이스에 권한 오류가 발생

 

문제 진단

[sylee@cray ~]$ ls -l /dev/kfd  crw-rw----. 1 root video 242, 0 Jun 10 14:33 /dev/kfd [sylee@cray ~]$ getent group video video:x:39: [sylee@cray ~]$ id uid=5003(sylee) gid=5003(sylee) groups=5003(sylee) context=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023

- /dev/kfd의 그룹은 video이고, 테스트 시 사용한 계정 sylee는 video 그룹에 등록되어져 있지 않다.

 

video 그룹에 sylee 계정 추가

[sylee@cray ~]$ sudo usermod -aG video sylee

 

video 그룹 계정 추가 확인

[sylee@cray ~]$ getent group video video:x:39:sylee [sylee@cray ~]$ id uid=5003(sylee) gid=5003(sylee) groups=5003(sylee),39(video) context=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023

 

rocminfo 재 실행

[sylee@cray ~]$ rocminfo  ROCk module is loaded =====================     HSA System Attributes     =====================     Runtime Version:         1.1 System Timestamp Freq.:  1000.000000MHz Sig. Max Wait Duration:  18446744073709551615 (0xFFFFFFFFFFFFFFFF) (timestamp count) Machine Model:           LARGE                               System Endianness:       LITTLE                              Mwaitx:                  DISABLED DMAbuf Support:          YES ==========                HSA Agents                ==========                *******                   Agent 1                   *******                     Name:                    Intel(R) Xeon(R) CPU E5-2670 v3 @ 2.30GHz   Uuid:                    CPU-XX                                Marketing Name:          Intel(R) Xeon(R) CPU E5-2670 v3 @ 2.30GHz   Vendor Name:             CPU                                   Feature:                 None specified                        Profile:                 FULL_PROFILE                          Float Round Mode:        NEAR                                  Max Queue Number:        0(0x0)                                Queue Min Size:          0(0x0)                                Queue Max Size:          0(0x0)                                Queue Type:              MULTI                                 Node:                    0                                     Device Type:             CPU                                   Cache Info:                   L1:                      32768(0x8000) KB      - - - - 이 하 생 략 - - - -