Bug: get_amd_offload_arch_flag in llamafile/cuda.c hardcodes "hipInfo" in some log messages

[tetron]

Contact Details

[email protected]

What happened?

In get_amd_offload_arch_flag it searches for both hipInfo and rocminfo, and uses the first one it finds. However, the log messages in this function only mention hipInfo which is misleading when one is trying to debug why GPU support isn't being built.

Version

llamafile 0.9.0

What operating system are you seeing the problem on?

Linux

Relevant log output

rocminfo is the one that returned non-zero exit, but the log messages make it seem like the problem is that hipInfo is the program that's missing or broken. This sent me down a debugging dead end.

2025-01-09T21:12:42.496150707Z get_rocm_bin_path: note: hipInfo not found on $PATH
2025-01-09T21:12:42.496152791Z get_rocm_bin_path: note: $HIP_PATH/bin/hipInfo does not exist
2025-01-09T21:12:42.496154254Z get_rocm_bin_path: note: /opt/rocm/bin/hipInfo does not exist
2025-01-09T21:12:42.496172849Z llamafile_log_command: /opt/rocm/bin/rocminfo
2025-01-09T21:12:42.523078724Z get_amd_offload_arch_flag: error: hipInfo returned non-zero exit status