gam-sae 0.3.155

//! Fixed-K sparse, minibatched SAE trainer (#1026, "collapsed linear lane").
//!
//! This is an **additive, standalone** path that makes very large dictionaries
//! (`K` up to tens of thousands) tractable, where the exact-REML / Arrow-Schur
//! dense joint manifold solver in [`crate::manifold`] is the wrong
//! engine: that solver carries a dense `N×K` latent state, `N×K×P` sensitivity
//! tensors, `K²N` penalty couplings, and a joint Newton over all `K` outer
//! parameters. None of that survives `K ≈ 32_000`.
//!
//! The collapsed linear lane instead trains a dictionary by alternating
//! minimisation with **no dense `N×K` object anywhere**:
//!
//! 1. **route** — for each row, score it against the whole dictionary in
//!    `K`-tiles (`scoring`) and keep only the top-`s` atoms online, so the
//!    `N×K` score matrix is produced one tile at a time and discarded;
//! 2. **codes** — solve the small `s×s` active-set least-squares system per row
//!    (`codes`), giving a fixed-width sparse code `(indices, codes)`;
//! 3. **decoder** — accumulate the sparse normal equations (method-of-optimal
//!    -directions / sparse GEMM) and refresh each atom (`update`);
//! 4. **project** — re-unit-norm every atom so the code scale is identified.
//!
//! All heavy state is FP32. The only dense `K`-sized objects are the decoder
//! itself (`K×P`) and the per-atom `P×P`/scalar accumulators — never `N×K`.
//!
//! The exact manifold engine is **untouched**: it remains the certification /
//! small-`K` path. This module is reached only through its own public entry
//! [`fit_sparse_dictionary`] (and the `gamfit` Python facade that wraps it).

mod block;
mod block_chart;
mod block_frame;
mod block_scoring_gpu;
mod block_stream;
mod code_evidence;
mod codes;
mod cofit;
mod coordinate;
#[cfg(target_os = "linux")]
mod decoder_gpu;
mod residual_reservoir;
#[cfg(target_os = "linux")]
mod score_router_backend;
mod scoring;
mod single_atom;
#[cfg(target_os = "linux")]
mod scoring_gpu;
mod split_lr_fdr;
mod stream;
mod update;

#[cfg(test)]
mod tests;

pub use block::{BlockSeedPolicy, BlockSparseConfig, BlockSparseConvergence, BlockSparseFit, BlockSparseFitError, block_gates, block_projections_row, block_sparse_dictionary_block_coords, block_sparse_dictionary_lift_block, block_sparse_dictionary_project_residual, block_sparse_dictionary_transform, coordinate_partition_frames, fit_block_sparse_dictionary, fit_block_sparse_dictionary_with_seed, reconstruct_block_sparse_rows, route_row_blocks};
pub use block_chart::{BlockChartComposeConfig, BlockChartComposeResult, BlockChartRecord, BlockSeedManifest, BlockSeedManifestConfig, BlockSeedRecord, CHART_FDR_ALPHA, ChartEvidence, MdlFeaturizerRow, block_sparse_dictionary_firings, block_sparse_dictionary_seed_manifest, compose_block_coordinate_charts};
pub use block_scoring_gpu::{BlockRoutePath, block_gate_row_cpu, route_blocks_cpu};
#[cfg(target_os = "linux")]
pub use block_scoring_gpu::{DEVICE_BLOCK_GATE_MIN_ELEMS, route_blocks_required};
pub use block_stream::{BlockEpochStats, BlockShardStats, BlockSparseStreamArtifact, BlockSparseStreamState};
pub use codes::SparseCode;
pub use cofit::{CofitConfig, CofitReport, CofitRound};
pub use coordinate::{BlockCoordinateReport, BlockMeasureCoordinateReport, FiringCoordinate, MeasureSpikeCoordinate, MeasureValuedCode, block_route_firing_coordinates, harmonic_measure_coordinates, harmonic_route_firing_coordinates, recover_measure_from_code};
pub use scoring::{ScoreRoutePath, ScoreRouteResult, ScoreRouteStats, TileScorer, top_s_online};
#[cfg(target_os = "linux")]
pub use scoring_gpu::{DEVICE_SCORE_BLOCK_MIN_ELEMS, ScoreBlockPath};
pub use split_lr_fdr::{
    FdrCertificate, crossfit_ui_log_evalue, family_fdr_certificate, shell_vs_ring_log_evalue,
};
pub use stream::{EpochStats, ShardStats, SparseDictArtifact, SparseDictStreamState};
pub use update::{DecoderSolveStats, LinearBlockRemlStats, SparseDictionaryError, linear_shared_rho_fs_step};
pub(crate) use update::extend_linear_reml_schedule;

use ndarray::{Array2, ArrayView2};

/// Shared (NOT per-atom) hyper-parameters for the collapsed linear lane.
///
/// The whole point of the sparse trainer is that `K` is too large to carry a
/// per-atom smoothing parameter / Newton state; every knob here is a single
/// scalar shared across the entire dictionary.
#[derive(Clone, Copy, Debug)]
pub struct SparseDictConfig {
    /// Dictionary width `K` (number of atoms).
    pub n_atoms: usize,
    /// Active budget `s`: how many atoms may fire per row (`top_s`). This is the
    /// shared routing-sparsity hyper-parameter.
    pub active: usize,
    /// Minibatch size (rows per route→code→accumulate step). The decoder is
    /// refreshed once per full epoch from the accumulated sparse normal
    /// equations, so this only bounds peak working set, not the solution.
    pub minibatch: usize,
    /// Number of full passes over the data.
    pub max_epochs: usize,
    /// Column tile width used when scoring rows against the dictionary. Score
    /// tiles of shape `minibatch × tile` are formed and discarded; the `N×K`
    /// score matrix is never materialised.
    pub score_tile: usize,
    /// Initial shared ridge on the per-row active-set code solve. The one-shot
    /// fit selects its final value by REML and records it in `selected_rho`.
    pub code_ridge: f32,
    /// Shared starting ridge; must equal `code_ridge` for a one-shot fit.
    /// Coupled MOD solves use it in their normal equations; a single-atom
    /// direction is profiled on its unit sphere.
    pub decoder_ridge: f32,
    /// Relative explained-variance improvement below which training stops.
    pub tolerance: f64,
    /// Per-fit score routing residency contract. `Required` is fail-closed: a
    /// high-`K` route that cannot run on the CUDA score-block path returns an
    /// error instead of silently scoring on the CPU.
    pub score_mode: gam_gpu::GpuPolicy,
}

impl SparseDictConfig {
    /// Construct a config for a `K`-atom dictionary, leaving every other knob at
    /// its shared default.
    pub fn new(n_atoms: usize) -> Self {
        Self {
            n_atoms,
            ..Self::default()
        }
    }
}

impl Default for SparseDictConfig {
    fn default() -> Self {
        Self {
            n_atoms: 1,
            active: 1,
            minibatch: 512,
            max_epochs: 30,
            score_tile: 4096,
            code_ridge: 1.0e-6,
            decoder_ridge: 1.0e-6,
            tolerance: 1.0e-6,
            score_mode: gam_gpu::GpuPolicy::Auto,
        }
    }
}

/// Result of a collapsed-linear-lane fit.
///
/// The routing is stored fixed-width and **sparse**: `indices[N, s]` /
/// `codes[N, s]`. There is deliberately no dense `N×K` assignment matrix —
/// reconstructing it would defeat the purpose of the lane.
#[derive(Clone, Debug)]
pub struct SparseDictFit {
    /// Decoder, `K×P`, unit-norm rows (one atom per row).
    pub decoder: Array2<f32>,
    /// Active atom indices per row, `N×s` (column `j` of row `i` is the `j`-th
    /// active atom for that row). Rows with fewer than `s` live atoms pad with
    /// repeated indices whose matching code is zero.
    pub indices: Array2<u32>,
    /// Sparse codes per row, `N×s`, aligned with [`Self::indices`].
    pub codes: Array2<f32>,
    /// Final held-in explained variance (`1 − RSS/TSS`).
    pub explained_variance: f64,
    /// Number of epochs actually run.
    pub epochs: usize,
    /// Inner-alternation and outer-REML fixed-point certificate. A
    /// [`SparseDictFit`] is constructed only when every residual below is at or
    /// below its matching tolerance.
    pub convergence: SparseDictConvergence,
    /// Active budget `s` actually used (`min(active, K)`).
    pub active: usize,
    /// Aggregate CPU/GPU scoring counters over every route pass in the fit.
    pub score_route_stats: ScoreRouteStats,
    /// Decoder refresh diagnostics from the final conditional direction update.
    pub decoder_solve_stats: DecoderSolveStats,
}

/// Checkable convergence evidence attached to every [`SparseDictFit`].
#[derive(Clone, Copy, Debug)]
pub struct SparseDictConvergence {
    /// Absolute held-in EV change over the final inner alternation.
    pub inner_ev_residual: f64,
    /// Configured inner EV tolerance.
    pub inner_tolerance: f64,
    /// Gauge-invariant decoder displacement under one full inner update map.
    pub decoder_residual: f64,
    /// Full-map, gauge-invariant decoder fixed-point threshold.
    pub decoder_tolerance: f64,
    /// Relative sparse-code/reconstruction displacement under one full inner map.
    pub routing_residual: f64,
    /// Full-map routing fixed-point threshold.
    pub routing_tolerance: f64,
    /// REML fixed-point residual `|ρ_new - ρ| / ρ`.
    pub outer_rho_residual: f64,
    /// Estimator-derived REML fixed-point tolerance.
    pub outer_tolerance: f64,
    /// Evidence-selected shared ridge.
    pub selected_rho: f64,
    /// Full inner fits evaluated by the outer REML schedule.
    pub outer_iterations: usize,
    /// Inner fits that paid the `O(K·N·P)` farthest-point seed. The schedule
    /// seeds exactly once and continues every later ρ from the prior iterate
    /// (#2441), so a schedule fit reports `1`; a hand-minted fit reports `0`.
    pub seeded_inner_runs: usize,
    /// Inner fits that continued from the previous iterate instead of seeding.
    /// Together with [`Self::seeded_inner_runs`] this sums to
    /// [`Self::outer_iterations`] for a schedule fit.
    pub continued_inner_runs: usize,
    /// Residual-row birth proposals that fired in the final inner transition.
    ///
    /// A positive count is compatible with an open certificate only when
    /// [`Self::support_saturated`] is true: those births replace live atoms on a
    /// fixed-cardinality support manifold instead of expanding model structure.
    pub accepted_births: usize,
    /// Largest live-atom cardinality reached during the final inner fit.
    pub live_atom_high_water: usize,
    /// Whether live-support cardinality set no new high for the full saturation
    /// confirmation window. This is reported independently of the EV plateau;
    /// both are required to return an open fit while births keep swapping (#2400).
    pub support_saturated: bool,
    /// Whether the inner fit reached the ABSOLUTE fixed point (EV, decoder AND
    /// routing residuals all within tolerance). `false` marks a **best-effort**
    /// fit returned at `K` above the intrinsic rank, where the `>rank` spurious
    /// support directions rotate freely in the equivalent-optima manifold and the
    /// routing residual legitimately cannot close (#2275) — the objective (EV) has
    /// plateaued but the discrete routing keeps churning. Convergence is decided by
    /// the gauge-invariant EV plateau, so both certified and open fits are returned;
    /// only a still-climbing objective (or a failed linear subsolve) is a genuine
    /// non-convergence error. Mirrors
    /// `super::block::BlockSparseConvergence::certified`.
    pub certified: bool,
}

impl SparseDictFit {
    /// Dense reconstruction `N×P` of the training rows from the sparse routing.
    ///
    /// This *does* allocate an `N×P` array (the data size, not `N×K`); it exists
    /// for diagnostics / EV checks, not as part of the trainer's hot loop.
    pub fn reconstruct(&self) -> Array2<f32> {
        reconstruct_sparse_rows(self.decoder.view(), self.indices.view(), self.codes.view())
            .expect("SparseDictFit stores internally validated routing")
    }
}

pub fn reconstruct_sparse_rows(
    decoder: ArrayView2<'_, f32>,
    indices: ArrayView2<'_, u32>,
    codes: ArrayView2<'_, f32>,
) -> Result<Array2<f32>, String> {
    if indices.dim() != codes.dim() {
        return Err(format!(
            "reconstruct_sparse_rows: indices shape {:?} does not match codes shape {:?}",
            indices.dim(),
            codes.dim()
        ));
    }
    let n = indices.nrows();
    let p = decoder.ncols();
    let mut out = Array2::<f32>::zeros((n, p));
    for i in 0..n {
        for j in 0..indices.ncols() {
            let atom = indices[[i, j]] as usize;
            if atom >= decoder.nrows() {
                return Err(format!(
                    "reconstruct_sparse_rows: atom index {atom} out of range 0..{}",
                    decoder.nrows()
                ));
            }
            let code = codes[[i, j]];
            if code == 0.0 {
                continue;
            }
            let row = decoder.row(atom);
            for c in 0..p {
                out[[i, c]] += code * row[c];
            }
        }
    }
    Ok(out)
}

/// Out-of-sample sparse-dictionary encode plus route-dispatch diagnostics.
#[derive(Clone, Debug)]
pub struct SparseDictTransform {
    /// Active atom indices per row, `M×active`.
    pub indices: Array2<u32>,
    /// Sparse codes per row, `M×active`.
    pub codes: Array2<f32>,
    /// CPU/GPU scoring counters for this transform route.
    pub score_route_stats: ScoreRouteStats,
}

/// Out-of-sample encode with an explicit score routing mode and route counters.
///
/// This is the Rust-native high-`K` T1 transform surface: callers that require
/// GPU scoring pass [`gam_gpu::GpuPolicy::Required`] and inspect
/// [`SparseDictTransform::score_route_stats`] to verify device engagement.
pub fn sparse_dictionary_transform_with_mode(
    x: ArrayView2<'_, f32>,
    decoder: ArrayView2<'_, f32>,
    active: usize,
    score_tile: usize,
    code_ridge: f32,
    score_mode: gam_gpu::GpuPolicy,
) -> Result<SparseDictTransform, String> {
    let k = decoder.nrows();
    if k == 0 {
        return Err("sparse_dictionary_transform: dictionary has no atoms".to_string());
    }
    if x.ncols() != decoder.ncols() {
        return Err(format!(
            "sparse_dictionary_transform: X has P={} columns but the decoder has P={}",
            x.ncols(),
            decoder.ncols()
        ));
    }
    let s = active.min(k).max(1);
    let scorer = TileScorer::new(s, score_tile.max(1));
    let routed = scorer.route_minibatch_with_mode(x, decoder, score_mode)?;
    let mut score_route_stats = ScoreRouteStats::default();
    score_route_stats.record_result(&routed);
    let m = x.nrows();
    let mut indices = Array2::<u32>::zeros((m, s));
    let mut codes = Array2::<f32>::zeros((m, s));
    for (row_idx, active_pairs) in routed.selections.iter().enumerate() {
        let code = codes::solve_row_codes(x.row(row_idx), decoder, active_pairs, s, code_ridge);
        for j in 0..s {
            indices[[row_idx, j]] = code.indices[j];
            codes[[row_idx, j]] = code.codes[j];
        }
    }
    Ok(SparseDictTransform {
        indices,
        codes,
        score_route_stats,
    })
}

/// Fit a fixed-`K` sparse minibatched linear dictionary to `x` (`N×P`).
///
/// This is the public entry of the collapsed linear lane. It never forms a
/// dense `N×K` object: scoring is tiled, routing is fixed-width sparse, and the
/// decoder is refreshed from accumulated sparse normal equations.
///
/// The two ridge fields must name one shared starting value. That value is only
/// the warm start for the evidence-selected shared `ρ`; every public fit runs the
/// outer REML fixed-point schedule and returns only after both inner and outer
/// certificates are satisfied.
pub fn fit_sparse_dictionary(
    x: ArrayView2<'_, f32>,
    config: &SparseDictConfig,
) -> Result<SparseDictFit, SparseDictionaryError> {
    update::run_linear_reml_schedule(x, config)
}