Module qp_layer 

ADMM-based differentiable QP layer with warm-start and active-set backward.

Solves the QP:

min ½ xᵀQx + cᵀx s.t. A_eq x = b_eq (equality) G_ineq x ≤ h_ineq (inequality)

The forward pass uses an OSQP-style ADMM iteration:

x-update: (Q + ρ Cᵀ C)⁻¹ (ρ Cᵀ (z - u) - c) where C = [A_eq; G_ineq] z-update: projection onto {Ax=b} × {Gx ≤ h} u-update: u += C x - z

The backward pass uses KKT sensitivity on the active constraints.

Structs

QpLayer

An ADMM-based differentiable QP layer.

QpLayerConfig

Configuration for the ADMM-based QP layer.