Skip to content

scratching

scratching

ScratchSparseCrossAttention

Bases: ForkingAttention

attention_weights(q: jax.Array, k: jax.Array, mask: Optional[jax.Array] = None, **kwargs: Any) -> jax.Array

Shared projected-Q/K probabilities, with original-position causality.

ScratchHardCrossAttention

Bases: ScratchSparseCrossAttention

Copy one causal source residual; differentiate the selector as softmax.