Arch Linux

This is missing a lot of context:

- Which package are you talking about?
- When you tweaked the PKGBUILD, could you share it?
- What problem are you trying to solve?

Future low quality bug reports like this will most likely just be closed, so please show some effort!

Yes. I apologise.
1st report, and I though I was doing it directly on the package.

package: python-pytorch-opt-rocm
PKGBUILD:
added on line 275:
export PYTORCH_ROCM_ARCH="gfx90c"

The package was originally built for other GPU's, so this variables PYTORCH_ROCM_ARCH were few GPU but mine is not in it.
And since the building is huge and always breaks due to OOM, I though of asking just to also add that GPU to the list on the next build

The current value of PYTORCH_ROCM_ARCH [1] does not include any APUs [2]? Perhaps a comment could be added explaining the selection?

[1] https://gitlab.archlinux.org/archlinux/packaging/packages/python-pytorch/-/blob/79b2507736b38bf87cf47a2c3a3e4aafbf7f2cd0/PKGBUILD#L275
[2] https://llvm.org/docs/AMDGPUUsage.html

The specific APU I am asking for:
gfx90c

So when I install the package and try to run pythorch I get:
2023-09-11 16:26:05.415264: I tensorflow/core/common_runtime/gpu/gpu_device.cc:2015] Ignoring visible gpu device (device: 0, name: AMD Radeon Graphics, pci bus id: 0000:04:00.0) with AMDGPU version : gfx90c. The supported AMDGPU versions are gfx1030, gfx900, gfx906, gfx908, gfx90a, gfx940, gfx941, gfx942.


So I would guess that if the build also adds that version I would be able to run it.
But I cannot make the build as I get OOM when I tried. and That is why I come here to ask for a package, also built with that flag

It's not as simple as just adding the targets to the list. Since APUs are not supported by ROCm, any low level GPU library will crash when using pytorch. Try

env HSA_OVERRIDE_GFX_VERSION=9.0.0 python ...

instead. This forces the GPU libraries to follow the code paths of the closest supported ISA, gfx900.