Arm Support

[Jothsa]

Is there any way to get this to work on Arm devices (specifically a raspberry pi 4)? I have managed to install everything, but when I try to run any of the commands I receive this error.

38555 illegal hardware instruction  python3 -m waifu2x.web.webgen.gen

I'm using 64 bit raspbian, and am inside a virtual environment. My python is version 3.9.2

[nagadomi]

I am not familiar with Raspberry Pi.
My first thought, the reason for the error is that PyTorch uses AVX/AVX2.
I recently wrote steps to build pytorch without AVX in the following Dockerfile. (Note that this even disables CUDA)
https://github.com/nagadomi/nunif/blob/master/Dockerfiles/Dockerfile.cpu_noavx
At that time, I read various blog posts about building PyTorch for Raspberry Pi.

However, I just searched and found that PyTorch supports aarch64, which is described in the official PyTorch article.
https://pytorch.org/tutorials/intermediate/realtime_rpi.html

I am sure that one of the pip modules uses instructions that are not supported by that CPU, but I don't know which one.

I think you should check the following.

1. Is pytorch working well?
2. which line is causing the error

For 2, I configured the faulthandler to detect C-level errors but it doesn't seem to be working.

nunif/waifu2x/web/server.py

Lines 456 to 460 in c4ef47f

	import logging
	import faulthandler
	faulthandler.enable()

Maybe it is crashing at module import time before main function.

EDIT:
I have built a Raspberry Pi3 aarch64 env on qemu.
It is still under setup, I will check it out later.

[Jothsa]

Thanks! I’ll check it out, but I might not have time until this weekend. I’ll keep you posted

[nagadomi]

This error is a bug in the pytorch aarch64 binary.
Just importing torch crashes.

$ python
Python 3.9.16 (main, Mar 30 2023, 13:22:28) 
[GCC 8.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
Illegal instruction

pytorch developers seem to be working on this error.
pytorch/pytorch#97226

EDIT:
This error happens only with pytorch 2.0, I will check if it works with 1.13.1.

[nagadomi]

It worked by downgrading pytorch to 1.13.1.

downgrade torchvision (torchvision try to install pytorch 2.0, so downgrade it first)

pip3 install torchvision==0.14

downgrade torch

pip3 install torch==1.13.1

Then it should work as normal.

Note that once the above pytorch issue is fixed, downgrading is no longer necessary.

[Jothsa]

Thanks so much!