Enzyme.jl Automatic Differentiation Skill

Enzyme.jl provides LLVM-level automatic differentiation for Julia, enabling high-performance gradient computation for both CPU and GPU code.

Type Annotations

Type annotations control how arguments are treated during differentiation:

Annotation	Description	Usage
`Const(x)`	Constant, not differentiated	Parameters, hyperparameters
`Active(x)`	Scalar to differentiate (reverse mode only)	Scalar inputs
`Duplicated(x, ∂x)`	Mutable with shadow accumulator	Arrays, mutable structs
`DuplicatedNoNeed(x, ∂x)`	Like Duplicated, may skip primal	Performance optimization
`BatchDuplicated(x, ∂xs)`	Batched shadows (tuple)	Multiple derivatives at once
`MixedDuplicated(x, ∂x)`	Mixed active/duplicated data	Custom rules with mixed types

using Enzyme