prepare_hvp

Julia

prepare_hvp(input; wrt, model_name, forcing, build_kwargs, source_layout, eval_cse, eval_cse_min_nodes)

prepare_hvp(input; wrt = (:states, :states), model_name = nothing, forcing = nothing,
            build_kwargs = NamedTuple(), source_layout = nothing) -> HVPEvaluator

Build the second-order contraction once (jacobian_bands2 ): each second-order coefficient field becomes a state of the derived evaluation model (the same mechanism as prepare_jacobian , whose keywords and source_layout contract this shares), and a precomputed (field slot, row, col, col2) list drives the contraction. A second column outside its variable’s extent is a ghost read on the state and parameter axes and an error on the forcing axis, as for the first-order columns.

Source

Python

prepare_hvp(obj, model_name: str | None = None, wrt=('states', 'states'), forcing=None) -> HVPEvaluator

Build an HVPEvaluator once — second-order differentiation, column spaces and point placement — for repeated hvp calls (one per Rosenbrock step of an adjoint sweep).

Source

Rust

pub fn prepare_hvp(file: &EsmFile, model_name: Option<&str>, wrt: (Wrt, Wrt), forcing: Option<&[String]>) -> Result<HvpEvaluator, BandError>

Build an HvpEvaluator once for the axis pair wrt (first axis states or time, second states, parameters or forcing).

Source · Rust docs