Struct SaeRowLayout 

pub struct SaeRowLayout {
    pub active_atoms: Vec<Vec<usize>>,
    pub coord_starts: Vec<Vec<usize>>,
    pub coord_offsets_full: Vec<usize>,
    pub coord_dims: Vec<usize>,
}

Per-row active-set layout for sparse SAE assignment (any mode).

When the assignment is sparse — structurally (JumpReLU gate) or effectively (softmax / IBP-MAP at large K, where the assignment mass concentrates on a small support) — only a subset of K atoms are active per observation. The Arrow-Schur row block for observation i has dim q_active_i = |active_atoms_i| + Σ_{k ∈ active_i} d_k rather than q = assignment_dim + Σ_k d_k. This struct records which atoms are active per row and maps compressed block positions back to full-q positions so that apply_newton_step can unpack the compact delta_t from the solve.

For JumpReLU the active set is exactly the gated support (a_{n,k} ≠ 0), so the compact solve is identity to the dense solve. For IBP-MAP the active set is the union of a top-k_active_cap truncation and a magnitude cutoff on a_{n,k}; this is only enabled when K is large enough that the dense (m_total · p)² data Gram would not fit the host / device working-set budget, and the dropped atoms carry O(a_{n,k}²) curvature that is negligible by construction of the cutoff.

#1408: SOFTMAX engages this compact layout when an explicit top_k (softmax_active_cap) and/or the in-core memory budget bounds the active set — the AssignmentMode::Softmax arm of assemble_arrow_schur consults crate::manifold::SaeManifoldTerm::softmax_active_plan and, on Some((cap, cutoff)), builds the active set via Self::from_dense_weights. The full-K dense softmax layout is retained only when neither lever engages (no top_k, in-budget K). Folding softmax top_k into the compact solve required writing the active×active Gershgorin Loewner majorizer sub-block (#1419; the softmax entropy curvature is indefinite, so its raw diagonal cannot be used) AND contracting that SAME majorizer over the compact logit slots in the logdet ρ-trace (assignment_log_strength_hessian_trace) and the θ-adjoint, so value, log|H|, and Γ differentiate one operator on the compact support. That coordinated change is landed and FD-certified; the FFI’s after-the-fit top-k projection is then a no-op at the optimum.

Fields

active_atoms: Vec<Vec<usize>>

active_atoms[row] — sorted indices of active atoms for that row.

coord_starts: Vec<Vec<usize>>

For row i, active atom active_atoms[i][j] has its coord block starting at compressed position coord_starts[i][j].

coord_offsets_full: Vec<usize>

Full-q coordinate offset for atom k (length k_atoms).

coord_dims: Vec<usize>

Per-atom coordinate dimensions, indexed by atom index.

Implementations

impl SaeRowLayout

pub fn row_q_active(&self, row: usize) -> usize

Per-row compressed dim.

pub fn expand_row(&self, row: usize, delta_t_row: &[f64], out: &mut [f64])

Expand a compact delta_t row slice back into full-q, zeros for inactive.

Trait Implementations

impl Clone for SaeRowLayout

fn clone(&self) -> SaeRowLayout

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for SaeRowLayout

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.