Max Weiss

Results 4 comments of Max Weiss

[Issue]: rocmi-smi and rsmi_compute_process_info_by_pid_get() don't show values for processes that don't have access to all installed GPUs (Linux)

It looks like the function `amd::smi::GetProcessInfoForPID` returns too early when the files are missing in /sys/devices/virtual/kfd/kfd/proc/. But I'm not sure if changing this function to just ignore the missing files...

[Issue]: rocmi-smi and rsmi_compute_process_info_by_pid_get() don't show values for processes that don't have access to all installed GPUs (Linux)

Unfortunately, ROCm 6.1.0 has the same issue. The host has four GPUs, but we mount only one into the container: ``` $ docker run --rm -it -v /opt/work:/opt/work --device=/dev/dri/card1 --device=/dev/dri/renderD128...

Only 8 of 64 GPUs Are Fully Partitioned and Usable in Docker After CPX/NPS4

ROCm 6.4.1 seems to have issues with MI300A and MI300X GPUs. I've created a similar issue in the ROCm repo: https://github.com/ROCm/ROCm/issues/4759. Older ROCm versions work fine.

Only 8 of 64 GPUs Are Fully Partitioned and Usable in Docker After CPX/NPS4

I think this is just a display error/bug. Does `rocminfo` show the two GPUs? On our host, the rocm-smi output in the container looks similar to yours, but `rocminfo` and...