candding 0.1.0

Pure candle embeddings for Rust with no ONNX: dense, sparse, multi-vector and rerank models today, image models on the roadmap
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
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

use candle_core::DType;
use candle_core::Module;
use candle_core::Tensor;
use candle_nn::Embedding;
use candle_nn::VarBuilder;
use candle_nn::embedding;

use super::config::BertConfig;
use crate::error::Result;
use crate::layers::layer_norm::LayerNorm;

pub struct BertEmbeddings {
  word: Embedding,
  position: Embedding,
  token_type: Embedding,
  layer_norm: LayerNorm,
}

impl BertEmbeddings {
  pub fn load(config: &BertConfig, vb: VarBuilder) -> Result<Self> {
    Ok(Self {
      word: embedding(
        config.vocab_size,
        config.hidden_size,
        vb.pp("word_embeddings"),
      )?,
      position: embedding(
        config.max_position_embeddings,
        config.hidden_size,
        vb.pp("position_embeddings"),
      )?,
      token_type: embedding(
        config.type_vocab_size,
        config.hidden_size,
        vb.pp("token_type_embeddings"),
      )?,
      layer_norm: LayerNorm::load(
        vb.pp("LayerNorm"),
        config.hidden_size,
        config.layer_norm_eps,
      )?,
    })
  }

  pub(super) fn weights_dtype(&self) -> DType {
    self.word.embeddings().dtype()
  }

  /// The `(vocab, hidden)` word embeddings, reused by a tied sparse head instead of loading a second copy.
  pub(crate) fn word_embeddings(&self) -> &Tensor {
    self.word.embeddings()
  }

  /// Absolute positions `0..seq`; right padding keeps real tokens at their true positions.
  pub fn forward(&self, input_ids: &Tensor, token_type_ids: &Tensor) -> Result<Tensor> {
    let (_, seq) = input_ids.dims2()?;
    let positions = Tensor::arange(0u32, seq as u32, input_ids.device())?.unsqueeze(0)?;
    let summed = (self.word.forward(input_ids)? + self.token_type.forward(token_type_ids)?)?;
    let summed = summed.broadcast_add(&self.position.forward(&positions)?)?;
    Ok(self.layer_norm.forward(&summed)?)
  }
}