Lib.rs

› Science › Machine learning › burn-common
#deep-learning #web-gpu #machine-learning #machine-learning-data

no-std burn-dispatch

Backend dispatch for the Burn framework

by Nathaniel Simard, Guillaume Lagrange

2 releases

Uses new Rust 2024

0.21.0-pre.3 Apr 8, 2026
0.21.0-pre.2 Mar 2, 2026

#2917 in Machine learning

Download history 18/week @ 2026-02-25 209/week @ 2026-03-04 298/week @ 2026-03-11 421/week @ 2026-03-18 665/week @ 2026-03-25 495/week @ 2026-04-01 747/week @ 2026-04-08

2,369 downloads per month

MIT/Apache

2MB
34K SLoC

Burn multi-backend dispatch.

Available Backends

The dispatch backend supports the following variants, each enabled via cargo features:

Backend Feature Description
Cpu cpu Rust CPU backend (MLIR + LLVM)
Cuda cuda NVIDIA CUDA backend
Metal metal Apple Metal backend via wgpu (MSL)
Rocm rocm AMD ROCm backend
Vulkan vulkan Vulkan backend via wgpu (SPIR-V)
Wgpu webgpu WebGPU backend via wgpu (WGSL)
NdArray ndarray Pure Rust CPU backend using ndarray
LibTorch tch Libtorch backend via tch
Autodiff autodiff Autodiff-enabled backend (used in combination with any of the backends above)

Note: WGPU-based backends (metal, vulkan, webgpu) are mutually exclusive. All other backends can be combined freely.

WGPU Backend Exclusivity

The WGPU-based backends (metal, vulkan, webgpu) are mutually exclusive due to the current automatic compile, which can only select one target at a time.

Enable only one of these features in your Cargo.toml:

  • metal
  • vulkan
  • webgpu

If multiple WGPU features are enabled, the build script will emit a warning and disable all WGPU backends to prevent unintended behavior.

Burn Backend Dispatch

A multi-backend dispatch that forwards the tensor operations to the appropriate backend.

Dependencies

~82–125MB
~2M SLoC