boostr 0.2.0

ML framework built on numr - attention, quantization, model architectures
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

//! Whisper decoder: causal transformer with cross-attention to encoder hidden states.
//!
//! Weight naming matches the HuggingFace `WhisperForConditionalGeneration` layout:
//! - `decoder.embed_tokens.weight`
//! - `decoder.embed_positions.weight`
//! - `decoder.layers.{i}.self_attn.{q,k,v,out}_proj.{weight,bias}`
//! - `decoder.layers.{i}.self_attn_layer_norm.{weight,bias}`
//! - `decoder.layers.{i}.encoder_attn.{q,k,v,out}_proj.{weight,bias}`
//! - `decoder.layers.{i}.encoder_attn_layer_norm.{weight,bias}`
//! - `decoder.layers.{i}.{fc1,fc2}.{weight,bias}`
//! - `decoder.layers.{i}.final_layer_norm.{weight,bias}`
//! - `decoder.layer_norm.{weight,bias}`
//! - `proj_out.weight` is tied to `decoder.embed_tokens.weight`
//!
//! The decoder uses pre-norm everywhere. Self-attention is causal; cross-attention
//! has no mask (encoder outputs are fully visible).

mod cache;
mod decoder;
mod helpers;
mod layer;

pub use cache::{DecoderCache, DecoderLayerCache};
pub use decoder::WhisperDecoder;
pub use layer::WhisperDecoderLayer;