- Full mHC pre_block/post_block with Sinkhorn-Knopp normalization - Dynamic A_l (sigmoid), B_l (Birkhoff polytope), C_l (2*sigmoid) - Checkpoint: attn_hc.fn (24,28672) + base (24,) + scale (3,) - Two mHC blocks per layer: attn_hc + ffn_hc - Removed emergency RMSNorm — mHC handles normalization properly - X_l: (1, n_hc=4, H) residual state, init from embedding broadcast
19 KiB
19 KiB