search

ROCm / amdsmi

AMD SMI
https://rocm.docs.amd.com/projects/amdsmi/en/latest
MIT License
29 stars 12 forks source link

[Issue]: different/inconsistent python interfaces #11

Open IlyasMoutawwakil opened 4 months ago

IlyasMoutawwakil commented 4 months ago

Problem Description

The python interface functions are different depending on rocm version, is there a reason they're not updated for all rocm versions ? rocm5.7: devices_handles = amdsmi.amdsmi_get_processor_handles() rocm5.6: devices_handles = amdsmi.amdsmi_get_device_handles()

my problem is that I have to do things like this to support each version of rocm 😅 https://github.com/huggingface/optimum-benchmark/blob/main/optimum_benchmark/launchers/isolation_utils.py#L65-L111

Operating System

Ubuntu 22.04

CPU

AMD Epyc

GPU

AMD Instinct MI250

ROCm Version

ROCm 5.7.0, ROCm 5.6.0

ROCm Component

amdsmi

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

marifamd commented 4 months ago

Hi @IlyasMoutawwakil thank you for reaching out. You are correct, we did change this because amd-smi will be able to handle all amd processors (GPUs & CPUs) in the future.

amd-smi is still a realtively new project and I'm excited to see developers use the python library that we developed. FYI we are aware of the bugs in the amdsmi_get_gpu_process_list & amdsmi_get_gpu_process_info functions as we are redesigning it.