Skip to content

token_attention

token_attention

Token-labeled attention on one fixed task example, attached to its trainer.

TokenAttention(spec: ExecutionSpec, base: Optional[Node] = None)

Bases: ReorderingAnalysis[Any]

Inspect one ordinary-attention head at the configured layer.

TokenSelector(spec: ExecutionSpec, base: Optional[Node] = None)

Bases: TokenAttention

Inspect both fork selectors and, when present, the final SSCA merger.