qualia_core_db/inference/sparse_cache.rs
1use crate::solvers::linear_algebra::{ConstTensorContractor, Tensor3x3x3};
2use crate::solvers::{SolverConfig, SolverResult, SolverState};
3
4/// Orthogonal Matching Pursuit (OMP) mathematically decomposes dense KV Cache vectors
5/// against a pre-loaded symbolic lattice dictionary.
6pub struct SparseDictionaryCache {
7 /// The pre-computed symbolic lattice (e.g., CML dictionaries mapped from WordNet).
8 /// For zero-allocation constraints, we model this as a fixed-size tensor.
9 pub dictionary: Tensor3x3x3,
10}
11
12impl SparseDictionaryCache {
13 /// Creates `SparseDictionaryCache` with the given state.
14 pub fn new(dictionary: Tensor3x3x3) -> Self {
15 Self { dictionary }
16 }
17
18 /// Decomposes a dense KV cache block into a sparse representation using
19 /// Orthogonal Matching Pursuit (OMP) via constant tensor contraction.
20 pub fn compress_kv_block(&self, dense_block: &Tensor3x3x3) -> SolverResult<Tensor3x3x3> {
21 // We use ConstTensorContractor to project the dense block against the dictionary.
22 let contractor = ConstTensorContractor {
23 tensor_a: *dense_block,
24 tensor_b: self.dictionary,
25 result: Tensor3x3x3::zero(),
26 contraction_indices: [(0, 0), (1, 1), (2, 2)], // Trace/Inner product approximation
27 config: SolverConfig::default(),
28 solver_state: SolverState::default(),
29 };
30
31 // Perform the constant tensor contraction to find the sparse coefficients.
32 let sparse_coefficients = contractor
33 .tensor_a
34 .contract(&contractor.tensor_b, &contractor.contraction_indices);
35
36 // Apply a hard threshold to enforce sparsity (Top-K approximation).
37 let mut thresholded = Tensor3x3x3::zero();
38
39 let threshold = 1e-3; // Define sparsity cutoff
40
41 for i in 0..3 {
42 for j in 0..3 {
43 for k in 0..3 {
44 let val = sparse_coefficients.get(i, j, k);
45 if val.abs() > threshold {
46 thresholded.set(i, j, k, val);
47 } else {
48 thresholded.set(i, j, k, 0.0);
49 }
50 }
51 }
52 }
53
54 Ok(thresholded)
55 }
56}