Burn Intermediate Representation

Burn's description of tensors and tensor operations as data

Every backend operation has a serializable description here. OperationIr and its per-kind enums (FloatOperationIr, IntOperationIr, BaseOperationIr, ...) name an operation and the TensorIrs it reads and writes. GraphIr groups operations with explicit inputs and outputs, and CustomOpIr carries operations defined by backend extensions.

Describing work as data rather than as calls lets it be inspected, optimized and moved before it runs. It is the basis of:

• kernel fusion, in burn-fusion;
• remote execution, in burn-remote through burn-router;
• graph capture, in burn-capture.

A backend that implements BackendIr can execute operations received in this form. Applications do not use this crate directly.

Feature Flags

• std (default): standard library support. Without it the crate is no_std with alloc.
• tracing: instrument operations with the tracing crate.

Part of the Burn deep learning framework. See the Burn Book and the API documentation.