ARM64 support for pytorch-directml

[ms300]

Hi, is there any plan with Pytorch-directml for Windows on ARM or ARM64 Linux(WSL 2)?

Only x86/x64 whls on Pypi for now: https://pypi.org/project/torch-directml/#files. And binaries included so i can't compile it by myself.

It is an effective lib for ML on Snapdragon devices and it would be nice to have to develop on Windows DevKit 2023!

[Adele101]

Thank you for submitting this issue. While I can't provide a timeline for resolution as the moment, please know that your feedback is valuable to us. We will follow up once we can review this issue.

[hez2010]

Is there any update to this issue? I want to use pytorch-directml on Windows arm64 based Copilot+PC.

[mcornier]

+1, Microsoft, you're not ready with Windows 11 on ARM. Why?
The Surface Pro X should have opened the path to this kind of support, but it's like it didn't.
The result: no DirectML torch for ARM64/AArch64 Windows or WSL2 Linux.
No real Qualcomm support, poor SDK...
And the x86 versions don't seem to work either.

PC with NPU ? I should stay with a NVidia GPU :(

[hez2010]

A short update: I've been able to instantiate a DirectML device on Windows 11 on ARM since DirectML 1.15.0. But we still need a pytorch-directml package for Windows ARM64.

[joshjkim]

@hez2010 @mcornier We recently ramped up support for torch-directml and have many new features coming before the year closes. Support for ARM64 is high up in that list :)

[VOM7HC]

Hi @joshjkim is there any new update? I am also waiting it

[ashumish-QCOM]

Hi

I noticed that there is a pull request on GitHub that enables PyTorch for Windows on ARM64 (CPU only). You might want to try this as a temporary solution: Enable Windows ARM64 by iremyux · Pull Request #133088. However, please note that this is CPU-only, and performance may be an issue.

pytorch/pytorch#133088

Thankyou

[VOM7HC]

However, please note that this is CPU-only, and performance may be an issue.

Look like torch directML backend for ARM still not ready, so there isn't accelerate from GPU or NPU

https://pypi.org/project/torch-directml/