Files
foxhunt/crates
jgrusewski 85b5f15520 feat: magnitude branch forward/backward uses [h_s2; Q_dir] conditioning
Forward: branch 1 FC GEMM reads [B, SH2+3] from mag_concat_buf.
All 3 forward paths (online_raw, online_f32, target_raw) updated.
Concat uses previous step's direction logits (one-step lag, negligible).

Backward: branch 1 dW uses wider X, dX writes to d_mag_concat_buf.
strided_accumulate extracts first SH2 columns into d_h_s2 with beta=1.
d_Q_dir (last 3 columns) discarded — detached, no backprop to direction.

Fixes backward goff_b_b1fc offset: uses ah*(sh2+3) matching compute_param_sizes.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-14 22:48:22 +02:00
..