miopenStatusUnknownError (AMD)

[Ecstatify]

Describe the bug Any operation that involves utilizing the AMD GPU through ROCm does not work in the WebUI. Running the training operation directly using the command generated by the WebUI works fine.

Additional context Just had to reinstall everything recently, with the ROCm set up done from memory. Now, RVC no longer works at all. I remember encountering this problem on my first install a few months ago, but I no longer remember the solution.

To Reproduce Run any task that uses the GPU in the WebUI. See the error in Terminal.

Here's a recount of what I did, just in case if I did something wrong in the installation:

(re)Install Ubuntu 22.04
Download the AMDGPU installer (6.1.2)
amdgpu-install --no-dkms --usecase=rocm,hiplibsdk
sudo usermod -aG render $USERNAME
sudo usermod -aG video $USERNAME
reboot
cd into the RVC installation kept on a different drive
deleted old venv and made new one
source venv/bin/activate
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.1
pip install -r requirements/amd.txt
export ROCM_PATH=/opt/rocm
export HSA_OVERRIDE_GFX_VERSION=10.3.0
python web.py

Expected behavior Tasks started in the WebUI do not error out.

Screenshots Says it completed training in the WebUI, but nothing actually happened. Train.log is empty

Desktop (please complete the following information):

• OS and version: Ubuntu 22.04
• Python version: 3.10

log.txt

[fumiama]

Maybe a problem from ROCm itself. Anyone who also met this problem?

[TheTrustedComputer]

I also run RVC on Arch Linux with ROCm and two 5500 XTs, and I don't have the same problem. Consider directly downloading the official PyTorch wheel that has ROCm support instead of using AMD's installer.