Files
foxhunt/crates
jgrusewski 524d81a45c feat: GLU backward replaces ReLU mask in branch backward pass
d_h_bd → glu_backward → d_value + d_gate_pre. Two weight gradient
GEMMs per branch (value + gate). Upstream gradient sums both paths.
VSN backward DEFERRED (W_vsn init to identity, zero gradient for now).
Gradient flows to gate weights at indices 34-41.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-15 00:14:48 +02:00
..