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
63
64
65
66
67
68
69
70
71
72
73
74
75

use candle_core::DType;
use candle_core::Tensor;
use candle_nn::VarBuilder;

use super::config::XlmRobertaConfig;
use super::embeddings::XlmRobertaEmbeddings;
use crate::error::Result;
use crate::layers::mask::AttentionMask;
use crate::models::Encoder;
use crate::models::EncoderInput;
use crate::models::bert::layer::BertLayer;

pub struct XlmRobertaModel {
  embeddings: XlmRobertaEmbeddings,
  layers: Vec<BertLayer>,
  config: XlmRobertaConfig,
}

/// Manual because `XlmRobertaEmbeddings` and `BertLayer` don't derive `Debug`; tests need it for `unwrap_err`.
impl std::fmt::Debug for XlmRobertaModel {
  fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
    f.debug_struct("XlmRobertaModel").finish_non_exhaustive()
  }
}

impl XlmRobertaModel {
  /// Catalog checkpoints use bare BERT tensor names with no `roberta.` prefix, so there is no prefix probe.
  pub fn load(config: XlmRobertaConfig, vb: VarBuilder) -> Result<Self> {
    config.ensure_supported()?;
    let embeddings = XlmRobertaEmbeddings::load(&config, vb.pp("embeddings"))?;
    let layers = (0..config.bert.num_hidden_layers)
      .map(|i| BertLayer::load(&config.bert, vb.pp(format!("encoder.layer.{i}"))))
      .collect::<Result<Vec<_>>>()?;
    Ok(Self {
      embeddings,
      layers,
      config,
    })
  }
}

impl Encoder for XlmRobertaModel {
  fn forward(&self, input: &EncoderInput) -> Result<Tensor> {
    let mut hidden = self
      .embeddings
      .forward(input.input_ids, input.token_type_ids)?;
    let mask = AttentionMask::padding(
      input.attention_mask,
      input.lengths,
      self.config.bert.num_attention_heads,
      self.config.head_dim(),
      hidden.dtype(),
    )?;
    for layer in &self.layers {
      hidden = layer.forward(&hidden, &mask)?;
    }
    Ok(hidden)
  }

  fn hidden_size(&self) -> usize {
    self.config.bert.hidden_size
  }

  /// The sequence limit, not the table height (514 on e5-base would index past the table).
  fn max_position_embeddings(&self) -> usize {
    self.config.sequence_limit()
  }

  fn weights_dtype(&self) -> DType {
    self.embeddings.weights_dtype()
  }
}

#[cfg(test)]
mod tests;