Skip to content

carved

`carved` ¶

FLT over a carved topology.

architecture/carve/attn and architecture/carve/mlp list the blocks whose attention and MLP sublayers survive; a block with neither is a None slot. architecture/carve/skips lists residual skip edges [source, dest]: the mean of the source blocks' outputs is added to the stream before block dest runs.

`CarvedFLTLoopedGPT` ¶

Bases: FLTLoopedGPT

`topology() -> tuple[set[int], set[int], tuple[tuple[int, ...], ...]]` ¶

Surviving attention blocks, surviving MLP blocks, skip sources per block.