tiyn/wiki
1
0
Fork 0
mirror of https://github.com/tiyn/wiki.git synced 2026-03-09 16:44:46 +01:00
Code Issues Projects Releases Wiki Activity

python: added rocm pytorch guide

Browse Source
This commit is contained in:
Marten B. Kante
2026-03-04 01:25:58 +01:00
parent b6e77adf69
commit 0b7874156a
1 changed files with 67 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

76
wiki/programming_language/python.md
View File

@@ -174,21 +174,20 @@ This section addresses the [PyTorch module](https://pytorch.org/).
Pytorch is a machine learning resource which is often used for
[neural networks](/wiki/neural_network.md).
#### Setup Pytorch with Cuda for GPU usage
#### Setup Pytorch with CUDA for GPU usage
Please note that according to
[various sources](https://www.reddit.com/r/archlinux/comments/1nxipcu/nvidia_pascal/gpu_not_supporting_cuda_13_can_i)
Cuda 13 does not support nVidia Pascal GPUs.
In this case an earlier version of Cuda has to be used.
CUDA is also only available for Nvidia GPUs.
For AMD GPUs refer to [the ROCm section](#setup-pytorch-with-rocm-for-gpu-usage).
If you are on Arch Linux or a distribution based on it install
`python-pytorch-cuda` via `pacman -S python-pytorch-cuda`.
After that visit
[the official pytorch website](https://pytorch.org/get-started/locally/) and
install pytorch for your custom configuration.
[the official PyTorch website](https://pytorch.org/get-started/locally) and
install PyTorch for your custom configuration.
After that try to run the following python script:
After that try to run the following python script.
This should give back `True` if the setup was successful and the GPU is available.
```python
import torch
@@ -196,7 +195,66 @@ import torch
torch.cuda.is_available()
```
This should give back `True`.
Please note that according to
[various sources](https://www.reddit.com/r/archlinux/comments/1nxipcu/nvidia_pascal/gpu_not_supporting_cuda_13_can_i)
CUDA 13 does not support nVidia Pascal GPUs.
In this case an earlier version of CUDA has to be used.
#### Setup PyTorch with ROCm for GPU usage
ROCm is also only available for AMD GPUs.
For Nvidia GPUs refer to [the CUDA section](#setup-pytorch-with-cuda-for-gpu-usage).
For ROCm to work some system packages have to be installed.
For Linux refer to the
[official Linux guide](https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/install/installrad/native_linux/install-radeon.html).
For [Arch Linux](/wiki/linux/arch-linux.md) although it is not listed the required
[AUR packages](/wiki/linux/package_manager.md) include `rocm-core`, `rocminfo`, `roctracer` and
`rccl`.
For Windows refer to the
[official Windows guide](https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/install/installrad/windows/howto_windows.html).
After the systemwide setup the project environment can be setup.
It is recommended to use [virtual environments](#using-virtual-environments).
The easiest way to achieve ROCm support is by using [pip](#modules).
As explained in the
[official PyTorch guide](https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/install/installrad/native_linux/install-pytorch.html)
the following example can be used inside the environment to install all needed ROCm packages.
In this example the version `7.2` is installed.
Adjustments to the command may have to be done in case another version should be installed.
```sh
wget https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torch-2.9.1%2Brocm7.2.0.lw.git7e1940d4-cp312-cp312-linux_x86_64.whl
wget https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torchvision-0.24.0%2Brocm7.2.0.gitb919bd0c-cp312-cp312-linux_x86_64.whl
wget https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/triton-3.5.1%2Brocm7.2.0.gita272dfa8-cp312-cp312-linux_x86_64.whl
wget https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torchaudio-2.9.0%2Brocm7.2.0.gite3c6ee2b-cp312-cp312-linux_x86_64.whl
pip install \
torch-2.9.1+rocm7.2.0.lw.git7e1940d4-cp312-cp312-linux_x86_64.whl \
torchvision-0.24.0+rocm7.2.0.gitb919bd0c-cp312-cp312-linux_x86_64.whl \
torchaudio-2.9.0+rocm7.2.0.gite3c6ee2b-cp312-cp312-linux_x86_64.whl \
triton-3.5.1+rocm7.2.0.gita272dfa8-cp312-cp312-linux_x86_64.whl
```
If old versions of `torch`, `torchvision`, `torchaudio` or `triton` are installed inside the
environment they may need to be removed.
After this installation for some GPUs - especially integrated GPUs like the Radeon 660M - an
additional step has to be taken.
In this case the following global shell variable has to be set.
```sh
export HSA_OVERRIDE_GFX_VERSION=10.3.0
```
After that try to run the following python script.
Since ROCm uses a bridge to access CUDA it should give back `True` if the setup was successful and
the GPU is available.
```python
import torch
torch.cuda.is_available()
```
### Hailo