Crate fusion 

Burn Fusion

Kernel fusion as a backend decorator.

Fusion wraps a backend that implements FusionBackend and defers its operations instead of running each one immediately. Queued operations are grouped by OperationFusers into Optimizations, so several element-wise operations can run as a single kernel and intermediate tensors never reach device memory.

Applications do not use this crate directly. The CubeCL backends (CUDA, ROCm, Metal, Vulkan, WebGPU, wgpu and the CubeCL CPU runtime) enable it by default. Backend authors implement FusionBackend and FusionRuntime to opt in; custom lets backend extensions register their own fused operations.

Feature flags

• std (default): standard library support.
• memory-checks: check the fusion runtime for leaked tensors (for tests).
• tracing: instrument operations with the tracing crate.

Modules

client

Client module exposing types to communicate with the fusion server.

custom

Types used to define and register custom Fusion operations.

observer

Watching the fusion server register operations and run blocks of them.

stream

Stream module exposing all tensor operations that can be optimized.

Macros

binary_float_cmp_ops

binary_float_ops

binary_int_cmp_ops

binary_int_ops

memory_checks

Export memory checks tests.

reduce_ops

scalar_float2int_ops

scalar_float_cmp_ops

scalar_float_ops

scalar_int_cmp_ops

scalar_int_ops

unary_float_ops

unary_int_ops

Structs

FuserProperties

The properties of a fuser.

Fusion

Enable dynamic operation fusion on a backend that implements fusion backend.

FusionTensor

Tensor primitive for the fusion backend for all kind.

NoOp

A no-operation placeholder for the fusion backend.

UnfusedOp

An operation that isn’t fused.

Enums

ExecutionError

An error that can happen when syncing a device.

FuserStatus

The status of a fuser.

OperationRan

Whether the operation that consumed a tensor ran.

Traits

FusionBackend

Trait that allows an existing backend to specify graph optimizations using operation fuser.

FusionRuntime

Trait that defines a runtime that will benefits from fused operations.

NumOperations

The number of operations contained in the data structure.

OperationFuser

The fusion operation abstraction allows implementations to fuse many tensor operations into one, improving the performance of the backend.

Optimization

The optimization created from a fuser.

Functions

get_client

Get the client for the given device.

Type Aliases

Client

Client alias.

FusionDevice

Type alias for <R as FusionRuntime>::FusionDevice.

FusionHandle

Type alias for <R as FusionRuntime>::FusionHandle.