Skip to content

MSA re-entry (wt/msa, 2026-10-02)

Lane wt/msa (Qwen Sparse Attention, tech report §QSA): gates 1-3 GREEN on CPU.

  • gate1 (sparse KB=all ≡ dense, bit-near < 2e-5): ndarray vendor/dormouse-fused/crates/burn-msa/tests/gate1_dense_contract.rs 3/3, flex crates/dormouse-core/tests/msa_gate1_flex.rs 2/2.
  • gate2 (indices in range, top-kb = scores’ own top, tail always in mask): tests/gate2_indices.rs 2/2.
  • gate3 (KL distill of max-pooled dense into indexer, ≥90% top-1 teacher agreement): tests/gate3_distill.rs 1/1 (68.8 s, 1000 AdamW @ 3e-3, synthetic).
  • Core wiring: use_msa off-default, --msa path via config seam, probe counters msa/msa_distill, eval line prints them. msa_kb >= B refused loud in validation.

Open: fused gather kernel (stage b, GPU — card busy), indexer LR group in the trainer (still Group::Rest), no A/B yet. Do not cite as quality evidence.