📓burn

burn

Thanks to lu-zero that give me a good idea to start the Rust backend for LocalAI project with burn.

Source: Author

Which platforms we need to choice for gRPC server by using burn?

The default acceleration of backend of the LocalAI project can support:

• CPU

  • OpenBLAS

• GPU

  • CUDA/CuBLAS - Nvidia

  • Hipplas - AMD

  • ClBLAS - AMD/Intel

  • Metal - Apple Silicon

So, the requirement of in here is that we need to have a backend that can support CPU and GPU, and as more as support the acceleration of CPU or GPU.

Supproted platforms of burn

According to the Supported Platforms of burn. We can se burn-ndarray backend does not support GPU. And the mainstream of AI framework is based on GPU. So, it is not a good idea as a default backend.

And Burn torch backend is based on tch-rs crate, which offers a Rust interface to the PyTorch C++ API. And it supports:

• CPU

• CUDA

• MPS

• Vulkan

As you can see that all the GPU acceleration was already support by LocalAI project.

Burn WGPU Backend is using the wgpu, and it supports

• Vulkan

• Metal

• DX11/12

• OpenGL

• WebGPU

So, here we can see that the WGPU backend can support all the GPU acceleration that LocalAI project need except the CPU accelertion. However, I beleive WASM could have a feature. And more and more LLMs need GPU to get a better performance. Although we have Lora and QLora technologies to decrease the computing resources of using by LLMs. But, we still need to have a GPU acceleration for LLMs. And here are some issues about the WGPU backend:

• https://github.com/Gadersd/stable-diffusion-xl-burn/issues/2

• https://github.com/Gadersd/stable-diffusion-burn/issues/7

• https://github.com/Gadersd/stable-diffusion-burn/issues/5

• https://github.com/Gadersd/whisper-burn/issues/19

According to above, it would be a good choice that we choose the Burn torch backend as the default backend of LocalAI project.

Although we do not have a benchmark here to test the performance of the WGPU backend. But, we can implement it as first. And deal with in the future development cycles if we have a better backend. There are also some exmaples of using burn, please check the current project on Github.

Reference

• Xmind address
• Build of LocalAI