Module attention

burn::tensor::backend::ops

Module attention

Expand description

Module with attention operations.

Functions§

naive_attention: Computes softmax(QKᵗ / √d) · V using separate kernels. Serves as a fallback when FlashAttention is not used.