PyTorch and ROCm 5: What ROCm packages are required for PyTorch?

1

I’m hoping to use PyTorch with ROCm to speed up some SVD using an AMD GPU. I’m new to GPU computing, ROCm and PyTorch, and feel a bit lost.

I’m pretty sure I need ROCm >= 5.0 to support the 6800 RX GPU, which means the PyTorch Get Started Locally command doesn’t quite work for me. Further, I’d like to test on a laptop with a Vega 8 iGPU which some ROCm packages do not support (HID, I believe).

Hence, I need to install ROCm differently, and due to my OS, I can’t use the AMD script (PopOS 21.10), but go through the Using Package Manager on Ubuntu method, which is all fine.

In the very last step, I’m asked to install ROCm meta-packages—and here I’m unsure what I need for PyTorch. AMD has a neat list of ROCm meta-packages, but I don’t know which are necessary.

I’ve been trying to prod the URL from the PyTorch Get Started Locally command, but I didn’t get anything out of that.

So, in short: Which of the ROCm meta-packages should I install to make use of PyTorch?

Thank you for your time.

1 Like
2

As far as I am aware (which I admit isn’t a lot) we don’t yet have Torch packages for ROCm 5.x and RDNA2 cards are only supported with ROCm 5.0 and above.

Although standalone Torch packages only support 4.5.2 at the moment there is a docker package from AMD which supports ROCm 5 but I’ve not experimented at all with this: https://www.amd.com/en/technologies/infinity-hub/pytorch

3

Ah. I didn’t realize there were specific PyTorch packages for the different ROCm versions. I thought that if I had the necessary ROCm packages installed, PyTorch would “just work” with them. Thank you for pointing this out!

Then I will sit curb my enthusiasm and wait until the “Start Locally” page is updated for ROCm 5.0+