ARM64 support

[ProKn1fe]

Any plans to support arm64 devices? Like Radxa Rock-5B or OrangePI 5.
It in any cases throw this.

Unhandled exception. System.IO.FileNotFoundException: Could not load file or assembly 'TorchSharp, Version=0.98.1.0, Culture=neutral, PublicKeyToken=67b0e5e1da516a1c'. The system cannot find the file specified.

[NiklasGustafsson]

No current plans for those devices, but contributions are always welcomed -- the Mac/ARM64 is being contributed by folks outside the Microsoft team.

[ProKn1fe]

So basically what need to make buildable on arm to run?

• LibTorchSharp
• ???

[GeorgeS2019]

pytorch/pytorch#13130

As long as the PyTorch community is considering Arm64 an open issue, it will not be VERY VERY trivial for TorchSharp

[ProKn1fe]

Seems they already support arm targets as nvidia jetson
https://github.com/pytorch/pytorch#nvidia-jetson-platforms

[GeorgeS2019]

https://pytorch.org/get-started/locally/

This is where official ARM64 support is needed before TorchSharp (I assume) will remotely consider ARM64

[NiklasGustafsson]

So basically what need to make buildable on arm to run?

• LibTorchSharp
• ???

ARM64 on Windows or Linux?

Theoretically, we should just need the ARM64 libtorch binaries, and then a setup to run tests in Azure Pipelines for release purposes. It should be possible to bundle the CPU support in the same NuGet as the other platforms, but I haven't thought through how it would impact the CUDA packages.

[ProKn1fe]

ARM64 on Windows or Linux?

Linux

[GeorgeS2019]

@NiklasGustafsson
Make more sense with Arm64 on Windows.

[MovGP0]

Builds on ARM64 are currently not working. Getting the following error: Visual Studio 2015, 2017 or 2019 required

Visual Studio is currently not available for ARM64. Visual Studio 2022 v17.4+ is supported on ARM64.

So maybe we should consider this possibility in the build scripts.

[NiklasGustafsson]

Well, there's no official distribution of libtorch for Windows / ARM64 (unless the team working on it has failed to let me know), so I wouldn't expect you to get much farther, anyway.

[p3root]

any updates on this?

[cm4ker]

There are ARM64 builds available in nightly (pytorch/pytorch#106956 (comment)). I tried to build TorchSharp with them, but I encountered some issues.