candle_transformers/
quantized_nn.rs

1//! Utilities for quanitized network layers
2//!
3//! This module contains various implementations of standard neural network layers, modules and
4//! utilities including embedding, linear layers, and various normalization techniques.
5//! Most implementations provide quantized weights support.
6
7use crate::models::with_tracing::QMatMul;
8use crate::quantized_var_builder::VarBuilder;
9use candle::quantized::QTensor;
10use candle::{Module, Result, Tensor};
11
12#[derive(Debug, Clone)]
13pub struct Embedding {
14    inner: candle_nn::Embedding,
15    span: tracing::Span,
16}
17
18impl Embedding {
19    pub fn new(d1: usize, d2: usize, vb: VarBuilder) -> Result<Self> {
20        let embeddings = vb.get((d1, d2), "weight")?.dequantize(vb.device())?;
21        let inner = candle_nn::Embedding::new(embeddings, d2);
22        let span = tracing::span!(tracing::Level::TRACE, "embedding");
23        Ok(Self { inner, span })
24    }
25
26    pub fn embeddings(&self) -> &Tensor {
27        self.inner.embeddings()
28    }
29}
30
31impl Module for Embedding {
32    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
33        let _enter = self.span.enter();
34        self.inner.forward(xs)
35    }
36}
37
38#[derive(Debug, Clone)]
39pub struct Linear {
40    weight: QMatMul,
41    bias: Option<Tensor>,
42}
43
44impl Linear {
45    pub fn from_arc(weight: std::sync::Arc<QTensor>, bias: Option<Tensor>) -> Result<Self> {
46        let weight = QMatMul::from_weights(weight)?;
47        Ok(Self { weight, bias })
48    }
49
50    pub fn from_weights(weight: QMatMul, bias: Option<Tensor>) -> Self {
51        Self { weight, bias }
52    }
53}
54
55impl Module for Linear {
56    fn forward(&self, x: &Tensor) -> candle::Result<Tensor> {
57        let x = x.apply(&self.weight)?;
58        match &self.bias {
59            None => Ok(x),
60            Some(bias) => x.broadcast_add(bias),
61        }
62    }
63}
64
65pub fn linear_b(in_dim: usize, out_dim: usize, bias: bool, vb: VarBuilder) -> Result<Linear> {
66    let bias = if bias {
67        Some(vb.get(out_dim, "bias")?.dequantize(vb.device())?)
68    } else {
69        None
70    };
71    let weight = QMatMul::new(in_dim, out_dim, vb)?;
72    Ok(Linear { weight, bias })
73}
74
75pub fn linear(in_dim: usize, out_dim: usize, vb: VarBuilder) -> Result<Linear> {
76    let bias = vb.get(out_dim, "bias")?.dequantize(vb.device())?;
77    let weight = QMatMul::new(in_dim, out_dim, vb)?;
78    Ok(Linear {
79        weight,
80        bias: Some(bias),
81    })
82}
83
84pub fn layer_norm(size: usize, eps: f64, vb: VarBuilder) -> Result<candle_nn::LayerNorm> {
85    let weight = vb.get(size, "weight")?.dequantize(vb.device())?;
86    let bias = vb.get(size, "bias")?.dequantize(vb.device())?;
87    Ok(candle_nn::LayerNorm::new(weight, bias, eps))
88}
89
90pub fn layer_norm_no_bias(size: usize, eps: f64, vb: VarBuilder) -> Result<candle_nn::LayerNorm> {
91    let weight = vb.get(size, "weight")?.dequantize(vb.device())?;
92    Ok(candle_nn::LayerNorm::new_no_bias(weight, eps))
93}
94
95pub fn linear_no_bias(in_dim: usize, out_dim: usize, vb: VarBuilder) -> Result<Linear> {
96    let weight = QMatMul::new(in_dim, out_dim, vb)?;
97    Ok(Linear { weight, bias: None })
98}
99
100#[derive(Debug, Clone)]
101pub struct RmsNorm {
102    weight: Tensor,
103    eps: f64,
104    span: tracing::Span,
105}
106
107impl RmsNorm {
108    pub fn new(size: usize, eps: f64, vb: VarBuilder) -> Result<Self> {
109        let span = tracing::span!(tracing::Level::TRACE, "rms-norm");
110        let weight = vb.get(size, "weight")?.dequantize(vb.device())?;
111        Ok(Self { weight, eps, span })
112    }
113
114    pub fn from_qtensor(weight: QTensor, eps: f64) -> Result<Self> {
115        let span = tracing::span!(tracing::Level::TRACE, "rms-norm");
116        let weight = weight.dequantize(&weight.device())?;
117        Ok(Self { weight, eps, span })
118    }
119}
120
121impl Module for RmsNorm {
122    fn forward(&self, x: &Tensor) -> Result<Tensor> {
123        let _enter = self.span.enter();
124        candle_nn::ops::rms_norm(x, &self.weight, self.eps as f32)
125    }
126}
candle_transformers/quantized_nn.rs

candle_transformers/
quantized_nn.rs
