TheTrustedComputer

Here's my attempt to force the compilation of hipBLASLt for my 5500 XT, which uses the gfx1012 architecture. It seemed to hit a brick wall when creating the ExtOp libraries....

@AngryLoki Do you mean any supported GPU architecture? I built it for mine,``-DAMDGPU_TARGETS=gfx1012``, which resulted in the error I posted above. I also appreciate your clarification regarding PyTorch's hipBLASLt requirement....

Removing the deprecation warning is understandable to stay current, but I have to rely on an older build of PyTorch (2.0.0) with ROCm acceleration (5.2) to enable local training on...

> Hmm... I guess we should stick to it for now, the plans to remove the `torch.nn.utils.weight_norm` from torch are not that immediate. For now, we should use the workaround...

Alternatively, you can explicitly call ``torch.load`` with the argument ``weights_only=True``. This will be the default in future PyTorch releases.

Just a fair warning: the current Mangio CREPE GPU implementation is broken on AMD cards, at least on my 5500 XT in a Linux environment. It completely crashes my GPU...

Agreed. If anyone has the expertise (I don't) to optimize Mangio CREPE for AMD GPUs without sacrificing performance on NVIDIA, please do so. Your contributions will be greatly appreciated.

You may or may not be aware of this, but there's a fork of RVC called [Applio](https://github.com/IAHispano/Applio) that has a working implementation of Mangio CREPE on AMD GPUs. They simplified...

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...

@charleswg Thank you for your insights. It appears NVIDIA cards aren't affected and may only apply to AMD.