Crate burn_cubecl 

The Burn backend for every CubeCL runtime.

CubeBackend implements Burn’s tensor operations as CubeCL kernels, compiled just in time for the device they run on. CUDA, ROCm, Metal, Vulkan, WebGPU, wgpu and the CubeCL CPU runtime all share this one backend type: a tensor’s CubeDevice says which runtime it uses. Cube is the type dispatch uses, wrapped in burn_fusion::Fusion when the fusion feature is on.

Applications reach this backend through a burn feature such as cuda, wgpu or cpu and a Device constructor; the runtime crates (burn-cuda, burn-wgpu, burn-rocm, burn-cpu) are thin wrappers that select a runtime. Use this crate directly to write custom kernels: kernel and ops hold the building blocks, and cubecl is re-exported so kernels use the same CubeCL version.

Feature flags

• cuda, hip, wgpu, metal, vulkan, webgpu, cpu: compile in a CubeCL runtime.
• fusion: kernel fusion through burn-fusion.
• autotune: benchmark kernel variants at runtime and keep the fastest.
• fft: FFT kernels.
• template: launch hand-written, non-JIT kernels.
• tracing: instrument operations with the tracing crate.

Re-exports

pub use element::BoolElement;

pub use element::CubeElement;

pub use element::FloatElement;

pub use element::IntElement;

pub use cubecl;

Modules

element

Elements for JIT backend

fusionfusion

Module for interacting with fusion

kernel

Kernel module

ops

Utilities for implementing JIT kernels

tensor

Tensor module.

Structs

CubeBackend

Tensor backend that compiles just-in-time for whichever runtime its device names.

CubeGraph

A captured launch sequence, tagged with the device it was captured on.

CubeTuneId

ID used to identify a Just-in-Time environment.

Enums

CubeAutotuneKey

Key for all autotune-enabled operations

CubeDevice

The device a cube tensor lives on.

Type Aliases

Cubefusion

The tensor backend for every cubecl runtime, fusing operations across streams. See CubeBackend for the unfused type.