pytorch-aarch64

pip install torch -f https://torch.kmtea.eu/whl/stable.html

Add torchvision, torchaudio, torchtext, torchcsprng and other packages if needed.

Consider using prebuilt wheels to speed up installation: pip install torch -f https://torch.kmtea.eu/whl/stable.html -f https://ext.kmtea.eu/whl/stable.html

(For users in China, please use the CDN)

torch	torchvision	torchaudio	torchtext	torchcsprng	Status	python
master nightly	master nightly	master nightly	master nightly	master nightly		>=3.6
1.10.0	0.11.1 0.11.0	0.10.0	0.11.0			>=3.6
1.9.1	0.10.1	0.9.1	0.10.1			>=3.6
1.9.0[i]	0.10.0	0.9.0	0.10.0			>=3.6 [i]
1.8.1	0.9.1 [i]	0.8.1	0.9.1	0.2.1		>=3.6
1.8.0 [i]	0.9.0	0.8.0	0.9.0	0.2.0		>=3.6
1.7.1	0.8.2	0.7.2	0.8.1	0.1.4		>=3.6
1.7.0	0.8.1 0.8.0	0.7.0	0.8.0	0.1.3		>=3.6
1.6.0 [i]	0.7.0	0.6.0	0.7.0	0.1.2 0.1.1 0.1.0		>=3.6
1.5.1	0.6.1	0.5.1	0.6.0			>=3.5
1.5.0	0.6.0	0.5.0	0.6.0			>=3.5
1.4.1 1.4.0	0.5.0	0.4.0	0.5.0			==2.7, >=3.5, <=3.8
1.3.1	0.4.2					==2.7, >=3.5, <=3.7
1.3.0	0.4.1					==2.7, >=3.5, <=3.7
1.2.0	0.4.0					==2.7, >=3.5, <=3.7
1.1.0	0.3.0					==2.7, >=3.5, <=3.7
<=1.0.1	0.2.2					==2.7, >=3.5, <=3.7

Corresponding Versions

• Corresponding torch and torchvision versions
• Corresponding torch and torchaudio versions
• Corresponding torch and torchtext versions

FAQ

• Q: Does this run on Raspberry Pi?
  A: Yes, if the architecture of the SoC is aarch64. It should run on all ARMv8 chips.
  
  
• Q: Does this support CUDA / CUDNN?
  A: No. Check here for more information.
  
  
• Q: Does this run on Nvidia Jetson?
  A: Yes, but extremely slow. Each Nvidia Jetson boards contains an Nvidia GPU, but this project only build cpu wheels. To better make use of your hardware, build it yourself.
  
  

Difference From The Official Wheels

In most circumstances, it’s recommended to just use the official wheels, and it will also be installed via pip by default, even with -f.

The wheels here are compiled from source on a rpi 4b+, and are for codes that crashed on official wheels, because of some unsupported instructions are used.

Use the torch wheels here only if you encounter problems like #8.

About Python 3.10

By the time this change (v1.9.0) is committed, NONE of the stable version of Python 3.10.0, Numpy 1.21.0 (which adds Python 3.10 support), or PyTorch 1.9.0 for Python 3.10 has been released.

If any critical issue is found, I may rebuild the wheel after stable releases.

About PyTorch v1.8.0

• Starting from v1.8.0, the official wheels of PyTorch for aarch64 has finally released!
  • To use the official wheels, use this index link:
    https://torch.kmtea.eu/whl/pfml.html
    where pfml stands for prefer-manylinux here.
    manylinux wheels will be installed by default.
• torchvision wheels are built with FFmpeg support. For wheels without it, please install torchvision==0.9.0+slim

About PyTorch v1.6.0

A fatal bug is encountered and this patch is applied while building PyTorch v1.6.0. The patch has been merged into mainstream in later versions.

About torchvision v0.9.1

Starting from torchvision v0.9.1, manylinux wheels are officially provided via both its indexes and PyPI. However, since they do not contain necessary backends (< 1MB) and may require extra installations, this project will continue to build torchvision wheels.

RuntimeError while importing

If you see something like this when import torch:

RuntimeError: module compiled against API version 0xe but this version of numpy is 0xd

Please upgrade your numpy version: pip install -U numpy.

CUDA / CUDNN Support

Since the building environment (as below) does not contain an Nvidia GPU, the wheels could not be built with cuda support.

If you need it, please use an Nvidia Jetson board to run the building code.

Building Environment

Host: Raspberry Pi 4 Model B

SoC: BCM2711 (quad-core A53)

Architecture: ARMv8 / ARM64 / aarch64

OS: Debian Buster

GCC: v8.3.0

Virtualization: Docker

Performance

Test date: 2021-10-29

Script: bench.py

Less execution time is better

Note:

1. This test was done by using a same Cat or Dog model, to predict 10 random animal images (while same for each group).
2. The latest version of PyTorch was manually installed on all platforms, but driver and Python remained default.