Godot Game Engine TorchSharp Environment

[GeorgeS2019]

This is WIP to figure out how to configure TorchSharp environment in Godot .NET6/7 environment
If anyone has experience, please share.

[NiklasGustafsson]

Is this an issue or a discussion? What specific action are you looking for from the maintainers?

[GeorgeS2019]

I solve the problem. Create issue for future Godot users coming to TorchSharp

  <ItemGroup>
    <PackageReference Include="libtorch-cuda-11.7-win-x64" Version="2.0.1.1" />
    <PackageReference Include="TorchSharp" Version="0.100.3" />
  </ItemGroup>

[NiklasGustafsson]

As mentioned high up on the README page:

We recommend using one of the 'bundled' packages: TorchSharp-cpu, TorchSharp-cuda-windows, or TorchSharp-cuda-linux. They will pull in the right libtorch backends.

[GeorgeS2019]

@NiklasGustafsson
I tried, so as @ChengYen-Tang
It does not work with Godot. It will take some time to investigate

[michieal]

okay, this kinda begs the question -- Godot builds for different platforms. Which one does one use (the package) for something like that?

[NiklasGustafsson]

The question of platform is (unfortunately) tied to libtorch as well as TorchSharp.

libtorch supports Windows x64, Linux x64, Linux ARM64, MacOS x64, MacOS M1. You can supposedly build it from source for other platforms, per my understanding. Of these, we haven't tested TorchSharp with Linux ARM or MacOS M1 (because I don't have the hardware for either).

[GeorgeS2019]

@NiklasGustafsson
Should I write a formal request to ask Microsoft allocate these hardware for you ☺️

[NiklasGustafsson]

@NiklasGustafsson Should I write a formal request to ask Microsoft allocate these hardware for you ☺️

No, thanks! :-)

[GeorgeS2019]

@NiklasGustafsson
I recently wrote a long email to someone why TorchSharp is important to Deep AI for 3D game engine like Godot with MIT license and ranked 8th in the entire GitHub universe.

[michieal]

The question of platform is (unfortunately) tied to libtorch as well as TorchSharp.

libtorch supports Windows x64, Linux x64, Linux ARM64, MacOS x64, MacOS M1. You can supposedly build it from source for other platforms, per my understanding. Of these, we haven't tested TorchSharp with Linux ARM or MacOS M1 (because I don't have the hardware for either).

So, different projects for different build platforms.
And, I understand the whole "don't have the hardware" aspect.

[NiklasGustafsson]

Not quite -- the TorchSharp-cpu package pulls in Windows/Linux/MacOS x64, so for those, you don't have to do anything special. If you need a CUDA backend, it's different. They are huge (2-3GB), so we don't pull them in by default.

[michieal]

okay, cool.
And, I wouldn't do the CUDA one as it's not guaranteed that the end user will have an Nvidia card, or ROCm drivers that would support it. That, and depending on the user's video card to run something like that, with it having full graphics, is pushing it. So... I would just use the CPU one.

Also, I saw that llama.cpp (github project) has a PR (at the time I saw it) that made use of the nmap? c++ command that allows the model files to be mapped as memory space (like a dedicated swap file) and I was wondering if TorchSharp has something like that or could work with something created in c# to do that?

(Probably should have asked that on a different topic/issue. my bad)

[NiklasGustafsson]

Yeah, turn that into a feature request...

[michieal]

Yeah, turn that into a feature request...

#1120