current workflow failed on AMD MI300X with rocm 6.2 and pytorch 2.4

[gfursin]

Hi,
I finally found some time to try the SCC'24 tutorial to run SDXL on AMD MI300X - the workflow resolved all dependencies but failed in loadgen. I attached the CM logs and deps.
Did you see such error - any ideas what is happening?
Which PyTorch version did you try?
Thank you,
Grigori
error-rocm.txt
error-rocm-deps.txt

Extra ref: mlcommons/cm4mlops#300

[arjunsuresh]

I'm not sure of the exact torch version used as it was done by the AMD team. But the below error looks like a HIP driver installation issue - the driver installation and detection for AMD GPUs are not done in CM like we do for Nvidia GPUs as we don't have a test system.

 File "/persistent_storage/gfursin/cm/lib/python3.10/site-packages/torch/cuda/__init__.py", line 319, in _lazy_init
    torch._C._cuda_init()
RuntimeError: No HIP GPUs are available

[gfursin]

Sure. We will investigate it further internally. Thanks!

[gfursin]

Fixed it - AMD ROCm drivers were not installed properly ;) .