xformers icon indicating copy to clipboard operation
xformers copied to clipboard

Support for ARM64 Architecture

Open KumoLiu opened this issue 1 year ago • 2 comments

🚀 Feature

Hi Development Team,

We are currently working on ARM64 supporting for our project, and would greatly appreciate it if xFormers could also support the ARM64 architecture and provide a corresponding wheel in the following repository: https://download.pytorch.org/whl/xformers.

Request features: Support for ARM64 Architecture:

Enable and test the build process for the ARM64 architecture. Ensure compatibility and optimal performance on ARM64 systems.

Provision of ARM64 Wheel: Generate and provide a pre-built wheel for the ARM64 architecture. Host the ARM64 wheel on https://download.pytorch.org/whl/xformers alongside the existing wheels for other architectures.

KumoLiu avatar Jul 17 '24 16:07 KumoLiu

I'm guessing the issue is that xformers has custom-built CUDA-kernels, that you'd have to rewrite them from scratch for MacOS's Metal-Shader (MPS) system, rather than CUDA, for xformers to be useful on ARM64 machines.

I don't think it's just a matter of changing the build target for the wheels.

PaulFidika avatar Nov 11 '24 01:11 PaulFidika

This is unlikely to happen anytime soon, as we don't have the time and resources to debug and fix issues on ARM64, nor to duplicate all the build and test jobs in our CI.

We might accept fixes that resolve compatibility with ARM64 if they do not introduce too much complexity, but this would remain best-effort and community-driven.

lw avatar Nov 12 '24 13:11 lw

Hi there, I'm developing my project on GH200 system using arm64 too, any update now guys ?

tbaro19 avatar May 04 '25 06:05 tbaro19

+1

weathon avatar Sep 21 '25 06:09 weathon

+1

DrAlexLiu avatar Nov 04 '25 23:11 DrAlexLiu

+1. Would like to get this working on DGX Spark GB10 which is arm64.

tymat avatar Nov 07 '25 02:11 tymat