Fedora amdgpu-install

Framework Laptop 16
1

It seems that Fedora39 conflicts with the installation of
amdgpu-install-6.0.60002-1718218.22.noarch.rpm
with

Error: Transaction test error:
  file /usr/bin from install of amdgpu-install-6.0.60002-1718218.22.noarch conflicts with file from package filesystem-3.18-6.fc39.x86_64

So the question is: How can we make sure we can use ROCM with mobile AMD GPU in Framework 16 under Fedora39?

Examples in Python would be helpful.

2

You can use Setting up ROCm and PyTorch on Fedora: A Step-By-Step Guide | by Anvesh G. Jhuboo | Medium to use the native packages to Fedora too.

1 Like
3

Thank you Mario.

The article is nice while it misses some elements for cupy (trying to find resources to get this last piece working).

Thank you again

4

After following all the steps of the article, I do get amd gpu ids not found.

I have try to set all the environment variables I could find on the web to point to GPU ID#0 since the framework has the additional GPU in first. Still get the same when running pytorch rocm 6.0.

Any idea?

5

I didn’t set any env variables and it just works for me. Not sure…

6

So I’m getting somewhere.

Rocm 6.0 does not manage properly the hybrid cards in my machine (under Fedora 40), whereas Rocm5.7 does perfectly well.

Here are the details:

Projects/learning-python/pytorch via 🐍 v3.12.2
❯ python
Python 3.12.2 (main, Feb 21 2024, 00:00:00) [GCC 14.0.1 20240217 (Red Hat 14.0.1-0)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> print(torch.__version__)
2.2.1+rocm5.7
>>> torch.cuda.is_available()
True
>>> torch.cuda.device_count()
1
>>> torch.cuda.current_device()
0
>>> torch.cuda.device(0)
<torch.cuda.device object at 0x7f913c3b2360>
>>> torch.cuda.get_device_name(0)
'AMD Radeon RX 7700S'
>>> quit()

Projects/learning-python/pytorch via 🐍 v3.12.2 took 57s
❯ source .venv/bin/activate

Projects/learning-python/pytorch via 🐍 v3.12.1 (.venv)
❯ python
Python 3.12.1 | packaged by Anaconda, Inc. | (main, Jan 19 2024, 15:51:05) [GCC 11.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
amdgpu.ids: No such file or directory
amdgpu.ids: No such file or directory
>>> print(torch.__version__)
2.3.0.dev20240312+rocm6.0
>>> torch.cuda.is_available()
True
>>> torch.cuda.device_count()
1
>>> torch.cuda.device(0)
<torch.cuda.device object at 0x7f0ff2000830>
>>> torch.cuda.get_device_name(0)
'AMD Radeon Graphics'
>>> quit()

So I guess I have to stick to Rocm 5.7 for now.

7

Ah that might explain the difference. I tested on Fedora 39 after all. Please file a bug with upstream, this sounds like a ROCm bug from your findings.