Skip to main content

Module fast

Module fast 

Source
Expand description

Fast implementations of commonly used multi-op functions.

Enums§

ScaledDotProductAttentionMask
Mask modes for scaled dot product attention.

Functions§

layer_norm
Layer normalization.
layer_norm_deviceDeprecated
Compatibility shim for layer_norm.
rms_norm
Root Mean Square normalization (RMS norm).
rms_norm_deviceDeprecated
Compatibility shim for rms_norm.
rope
Optimized implementation of NN.RoPE.
rope_deviceDeprecated
Compatibility shim for rope.
rope_dynamic
Optimized implementation of NN.RoPE with dynamic (array) offset.
rope_dynamic_deviceDeprecated
Compatibility shim for rope_dynamic.
scaled_dot_product_attention
A fast implementation of multi-head attention: O = softmax(Q @ K.T, dim=-1) @ V
scaled_dot_product_attention_deviceDeprecated
Compatibility shim for scaled_dot_product_attention.