Struct PirlsStepStreamDeviceInput 

pub struct PirlsStepStreamDeviceInput<'a, 'b> {
    pub w_solver_dev: &'a CudaSlice<f64>,
    pub grad_eta_dev: &'b CudaSlice<f64>,
    pub penalty_hessian: ArrayView2<'b, f64>,
    pub step_lm_lambda: f64,
    pub objective_ridge: f64,
    pub beta_dev: &'b CudaSlice<f64>,
    pub linear_shift: ArrayView1<'b, f64>,
}

Stage 3.2 device-input variant of PirlsStepStreamInput.

Where the host-input form uploads weights + gradient per Newton step, this form reads them straight from the crate::gpu_kernels::pirls_row::RowOutputDevBuffers populated by the device-side row-reweight kernel — no host round-trip for the row state. Only the penalty matrix still crosses the host boundary because the outer REML loop updates Sλ + LM ridge between PIRLS steps.

Fields

w_solver_dev: &'a CudaSlice<f64>

Device-resident solver weights w_solver_i (length n). Read in-place by the cublasDdgmm WX assembly.

grad_eta_dev: &'b CudaSlice<f64>

Device-resident IRLS gradient ∂ℓ/∂η_i (length n). Read by the Xᵀg dgemv to form the Newton RHS.

penalty_hessian: ArrayView2<'b, f64>

Penalty Hessian Sλ in row-major host layout (p × p).

step_lm_lambda: f64

Temporary LM damping for this Newton solve step only. Added to H before potrf; stripped out of the snapshotted penalized_hessian.

objective_ridge: f64

Real model-objective ridge. Appears in the exported penalized_hessian that flows to EDF / REML curvature.

beta_dev: &'b CudaSlice<f64>

Current coefficient vector β (length p). Downloaded to the host to form the Newton RHS correction S·β. Only p f64 values cross the boundary (β is small), so the round-trip cost is negligible.

linear_shift: ArrayView1<'b, f64>

Linear shift vector (length p) in transformed coordinates, on host. Added to Newton RHS so the solve targets Xᵀ·score − S·β + linear_shift.