memorax.networks.layers#
Low-level layer primitives.
Convolution#
CausalConv1d - Stateful causal 1D convolution for recurrent use.
ParallelCausalConv1d - Parallel (non-recurrent) causal convolution.
Dense#
BlockDiagonalDense - Block-diagonal dense layer for efficient computation.
Normalization#
MultiHeadLayerNorm - Per-head layer normalization.
Utility#
Flatten - Reshape to batch × feature.
Identity - Pass-through layer.