liblaf-peach 0.10.0

Torch-first optimization protocols and CuPy linear-solver helpers for numerical experiments.

Explore the docs »

Documentation · Changelog · Issues

What It Is

Peach is a Torch-based toolbox for optimization and linear-solver experiments. It keeps problem definitions small: optimizers ask concrete problem objects for objective hooks, and linear solvers ask for matrix-vector hooks.

It contains:

• Protocol-based optimizer and linear-system interfaces.
• A preconditioned nonlinear conjugate-gradient optimizer with Armijo backtracking, adaptive diagonal Hessian damping, and optional problem hooks for callbacks and step-size limits.
• CuPy-backed conjugate-gradient and MINRES wrappers for torch tensors, with residual diagnostics.
• A SciPy optimizer adapter and a Rosenbrock problem for tests and examples.

Install

uv add liblaf-peach

Example

import torch

from liblaf.peach.optim.pncg import Pncg


class QuadraticProblem:
    def __init__(self, target):
        self.target = target

    def update(self, state, params, /):
        state.copy_(params)

    def fun(self, state, /):
        residual = state - self.target
        return 0.5 * torch.dot(residual, residual)

    def grad(self, state, /):
        return state - self.target

    def hess_diag(self, state, /):
        return torch.ones_like(state)

    def hess_quad(self, state, direction, /):
        return torch.dot(direction, direction)


params = torch.tensor([0.0])
model_state = params.clone()
problem = QuadraticProblem(target=torch.tensor([3.0]))
solution = Pncg().minimize(problem, model_state, params)

print(solution.params)
print(model_state)

Local Development

gh repo clone liblaf/peach
cd peach
mise run install
uv run pytest

License

liblaf-peach is licensed under the MIT License.