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
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120

use crate::models::with_tracing::QMatMul;
use crate::quantized_var_builder::VarBuilder;
use candle::quantized::QTensor;
use candle::{Module, Result, Tensor};

#[derive(Debug, Clone)]
pub struct Embedding {
    inner: candle_nn::Embedding,
    span: tracing::Span,
}

impl Embedding {
    pub fn new(d1: usize, d2: usize, vb: VarBuilder) -> Result<Self> {
        let embeddings = vb.get((d1, d2), "weight")?.dequantize(vb.device())?;
        let inner = candle_nn::Embedding::new(embeddings, d2);
        let span = tracing::span!(tracing::Level::TRACE, "embedding");
        Ok(Self { inner, span })
    }

    pub fn embeddings(&self) -> &Tensor {
        self.inner.embeddings()
    }
}

impl Module for Embedding {
    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
        let _enter = self.span.enter();
        self.inner.forward(xs)
    }
}

#[derive(Debug, Clone)]
pub struct Linear {
    weight: QMatMul,
    bias: Option<Tensor>,
}

impl Linear {
    pub fn from_arc(weight: std::sync::Arc<QTensor>, bias: Option<Tensor>) -> Result<Self> {
        let weight = QMatMul::from_weights(weight)?;
        Ok(Self { weight, bias })
    }

    pub fn from_weights(weight: QMatMul, bias: Option<Tensor>) -> Self {
        Self { weight, bias }
    }
}

impl Module for Linear {
    fn forward(&self, x: &Tensor) -> candle::Result<Tensor> {
        let x = x.apply(&self.weight)?;
        match &self.bias {
            None => Ok(x),
            Some(bias) => x.broadcast_add(bias),
        }
    }
}

pub fn linear_b(in_dim: usize, out_dim: usize, bias: bool, vb: VarBuilder) -> Result<Linear> {
    let bias = if bias {
        Some(vb.get(out_dim, "bias")?.dequantize(vb.device())?)
    } else {
        None
    };
    let weight = QMatMul::new(in_dim, out_dim, vb)?;
    Ok(Linear { weight, bias })
}

pub fn linear(in_dim: usize, out_dim: usize, vb: VarBuilder) -> Result<Linear> {
    let bias = vb.get(out_dim, "bias")?.dequantize(vb.device())?;
    let weight = QMatMul::new(in_dim, out_dim, vb)?;
    Ok(Linear {
        weight,
        bias: Some(bias),
    })
}

pub fn layer_norm(size: usize, eps: f64, vb: VarBuilder) -> Result<candle_nn::LayerNorm> {
    let weight = vb.get(size, "weight")?.dequantize(vb.device())?;
    let bias = vb.get(size, "bias")?.dequantize(vb.device())?;
    Ok(candle_nn::LayerNorm::new(weight, bias, eps))
}

pub fn layer_norm_no_bias(size: usize, eps: f64, vb: VarBuilder) -> Result<candle_nn::LayerNorm> {
    let weight = vb.get(size, "weight")?.dequantize(vb.device())?;
    Ok(candle_nn::LayerNorm::new_no_bias(weight, eps))
}

pub fn linear_no_bias(in_dim: usize, out_dim: usize, vb: VarBuilder) -> Result<Linear> {
    let weight = QMatMul::new(in_dim, out_dim, vb)?;
    Ok(Linear { weight, bias: None })
}

#[derive(Debug, Clone)]
pub struct RmsNorm {
    weight: Tensor,
    eps: f64,
    span: tracing::Span,
}

impl RmsNorm {
    pub fn new(size: usize, eps: f64, vb: VarBuilder) -> Result<Self> {
        let span = tracing::span!(tracing::Level::TRACE, "rms-norm");
        let weight = vb.get(size, "weight")?.dequantize(vb.device())?;
        Ok(Self { weight, eps, span })
    }

    pub fn from_qtensor(weight: QTensor, eps: f64) -> Result<Self> {
        let span = tracing::span!(tracing::Level::TRACE, "rms-norm");
        let weight = weight.dequantize(&weight.device())?;
        Ok(Self { weight, eps, span })
    }
}

impl Module for RmsNorm {
    fn forward(&self, x: &Tensor) -> Result<Tensor> {
        let _enter = self.span.enter();
        candle_nn::ops::rms_norm(x, &self.weight, self.eps as f32)
    }
}