| Git Clone URL: | https://aur.archlinux.org/python-torchvision-rocm.git (read-only, click to copy) |
|---|---|
| Package Base: | python-torchvision-rocm |
| Description: | Datasets, transforms, and models specific to computer vision (C++ library only, with ROCM support) |
| Upstream URL: | https://github.com/pytorch/vision |
| Licenses: | BSD-3-Clause |
| Conflicts: | torchvision |
| Provides: | torchvision |
| Submitter: | wuxxin |
| Maintainer: | wuxxin |
| Last Packager: | wuxxin |
| Votes: | 9 |
| Popularity: | 0.78 |
| First Submitted: | 2022-04-18 13:21 (UTC) |
| Last Updated: | 2024-11-15 14:31 (UTC) |
for AMD GPU Types i mirror the config of the corresponding pytorch-rocm package,
because if the main packages is not build (per default) for an GPU architecture,
it does not make sense to build torchvision for more architectures than the main package.
The currently supported AMD-GPU types of the python-pytorch-rocm* package is found under:
But you can always add either of $GPU_TARGETS, $AMDGPU_TARGETS or $PYTORCH_ROCM_ARCH
to a custom GPU type before building python-torchvision-rocm.
+1 for dreieck's request, and if you could please also add gfx90c we'd be grateful!
Can you please add further AMD GPUs to _PYTORCH_ROCM_ARCH?
Mine e.g. is gfx1103. See here for a list.
Just in case anyone encounters the same issue as me, I'll leave my solution here:
I was trying to install python-torchvision-rocm, however since I had opencl-amd installed, that provided the librocm_smi64.so binary. The version of this one is at librocm_smi64.so.5, while python-torchvision-rocm expects librocm_smi64.so.1, so it was throwing this error: ImportError: librocm_smi64.so.1: cannot open shared object file: No such file or directory.
To fix this, I simply added a symlink from librocm_smi64.so.1 to the currently installed binary (in my case librocm_smi64.so.5.0.50700 located in /opt/rocm/lib/) by running sudo ln -s librocm_smi64.so.5.0.50700 /opt/rocm/lib/librocm_smi64.so.1.
Probably not the cleanest approach and an update might break it, however it works for now. If anyone knows about a better way to fix this, feel free to reply.
It fails on AMD because there are 2 lines with FORCE_CUDA=1 in the PKGBUILD, which makes it only work on NVIDIA.
@TheJackiMonster are you trying use package with nvidia GPU? This package is for AMD ROCm. For cuda there is torchvision-cuda in official Arch repo
CUDA seems to be required to build the package: https://archlinux.org/packages/community/x86_64/cuda/
Also: "CUDA_HOME environment variable is not set. Please set it to your CUDA install root."
I would recommend adding cuda to the build requirements and set CUDA_HOME appropiately.
Development is on Github: https://github.com/wuxxin/aur-packages Please open issues and PRs there instead of commenting.
Pinned Comments
wuxxin commented on 2022-04-18 13:29 (UTC) (edited on 2024-03-20 18:27 (UTC) by wuxxin)
Development is on Github: https://github.com/wuxxin/aur-packages Please open issues and PRs there instead of commenting.