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
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161

use std::{collections::HashMap, ops::Mul, sync::Arc};

use candle_core::{Module, Result, Tensor};
use candle_nn::{init, Embedding, Init, VarBuilder};
use either::Either;

use crate::{
    frozenembed::FrozenEmbedding, EmbeddingLayerLike, LoraConfig, Merge, MergeError,
    MergeErrorOrError, Saveable,
};

#[derive(Debug, Clone)]
pub struct LoraEmbedding {
    old: Arc<FrozenEmbedding>,
    embed_a: Embedding,
    a: Tensor,
    b: Tensor,
    scale: Option<f64>,
    merged: bool,
    prefix: String,
    id: usize,
}

#[derive(Clone, Debug)]
/// Configuration for LoraEmbedding, with `num_embeddings` vectors of `embedding_dim` size`.
pub struct LoraEmbeddingConfig {
    num_embeddings: usize,
    embedding_dim: usize,
}

impl LoraEmbeddingConfig {
    pub fn new(num_embeddings: usize, embedding_dim: usize) -> Self {
        LoraEmbeddingConfig {
            num_embeddings,
            embedding_dim,
        }
    }
}

impl LoraEmbedding {
    pub fn new(
        old: &dyn EmbeddingLayerLike,
        embed_config: &LoraEmbeddingConfig,
        config: &LoraConfig,
        vb: &VarBuilder,
        id: usize,
    ) -> Result<Self> {
        let a = vb.pp(format!("a{id}")).get_with_hints(
            (config.rank, embed_config.num_embeddings),
            "weight",
            init::ZERO,
        )?;
        let b: Tensor = vb.pp(format!("b{id}")).get_with_hints(
            (embed_config.embedding_dim, config.rank),
            "weight",
            Init::Randn {
                mean: 0.0,
                stdev: 1.0,
            },
        )?;

        let mut a_t = a.t()?;
        a_t = a_t.reshape(a_t.shape())?;
        let embed_a = Embedding::new(a_t.clone(), a_t.dim(1)?);

        Ok(LoraEmbedding {
            old: Arc::new(FrozenEmbedding::new_from_embed(old)?),
            embed_a,
            a,
            b,
            scale: if config.rank > 0 {
                Some(config.alpha / config.rank as f64)
            } else {
                None
            },
            merged: false,
            prefix: vb.prefix(),
            id,
        })
    }
}

impl Merge for LoraEmbedding {
    fn get_delta_weight(&self) -> std::result::Result<Tensor, MergeErrorOrError> {
        let result = self.b.matmul(&self.a).map_err(Either::Right)?;
        Ok(match self.scale {
            Some(scale) => result.mul(scale).map_err(Either::Right)?,
            None => result,
        })
    }

    fn merge_weights(&mut self) -> std::result::Result<(), MergeErrorOrError> {
        if self.merged {
            Err(Either::Left(MergeError::AlreadyMerged))
        } else {
            self.old = Arc::new(
                FrozenEmbedding::new(
                    &(self.embeddings() + self.get_delta_weight()?.transpose(0, 1))
                        .map_err(Either::Right)?,
                    self.hidden_size(),
                )
                .map_err(Either::Right)?,
            );
            self.merged = true;
            Ok(())
        }
    }

    fn unmerge_weights(&mut self) -> std::result::Result<(), MergeErrorOrError> {
        if !self.merged {
            Err(Either::Left(MergeError::NotMerged))
        } else {
            self.old = Arc::new(
                FrozenEmbedding::new(
                    &(self.embeddings() - self.get_delta_weight()?.transpose(0, 1))
                        .map_err(Either::Right)?,
                    self.hidden_size(),
                )
                .map_err(Either::Right)?,
            );
            self.merged = false;
            Ok(())
        }
    }
}

impl Module for LoraEmbedding {
    fn forward(&self, input: &Tensor) -> Result<Tensor> {
        let mut result = self.old.forward(input)?;
        if let Some(scale) = self.scale {
            let b = self.b.t()?;
            let b = b.reshape(b.shape())?;

            let after_a = self.embed_a.forward(input)?;
            result = (result + (after_a.broadcast_matmul(&b)?).mul(scale))?
        }
        Ok(result)
    }
}

impl Saveable for LoraEmbedding {
    fn get_tensors(&self, accum: &mut HashMap<String, Tensor>) {
        accum.insert(
            self.prefix.clone() + &format!(".a{}.weight", self.id),
            self.a.clone(),
        );
        accum.insert(
            self.prefix.clone() + &format!(".b{}.weight", self.id),
            self.b.clone(),
        );
    }
}

impl EmbeddingLayerLike for LoraEmbedding {
    fn embeddings(&self) -> &Tensor {
        self.old.embeddings()
    }
    fn hidden_size(&self) -> usize {
        self.old.hidden_size()
    }
}