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
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
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234

use candle_core::Device;
use candle_nn::VarMap;
use tempfile::tempdir;

use super::*;
use crate::layers::activation::Activation;
use crate::models::bert::BertConfig;

fn tiny_config() -> BertConfig {
  BertConfig {
    vocab_size: 64,
    hidden_size: 1,
    num_hidden_layers: 0,
    num_attention_heads: 1,
    intermediate_size: 1,
    hidden_act: Activation::Gelu,
    max_position_embeddings: 32,
    type_vocab_size: 1,
    layer_norm_eps: 1e-12,
  }
}

/// `BAAI/bge-m3`'s own `pad_token_id` (1) and `position_embedding_type` ("absolute").
fn xlm_config(bert: BertConfig) -> XlmRobertaConfig {
  XlmRobertaConfig {
    bert,
    pad_token_id: 1,
    position_embedding_type: "absolute".to_string(),
  }
}

/// Never run: every `scatter_rows` test builds `hidden` by hand.
fn dummy_encoder() -> XlmRobertaModel {
  let varmap = VarMap::new();
  let vb = VarBuilder::from_varmap(&varmap, DType::F32, &Device::Cpu);
  XlmRobertaModel::load(xlm_config(tiny_config()), vb).unwrap()
}

fn head(weight: Tensor, bias: Tensor, vocab_size: usize) -> M3Sparse {
  M3Sparse {
    encoder: dummy_encoder(),
    weight,
    bias,
    vocab_size,
    dropped: DROPPED,
  }
}

/// Ids `[5, 5]` with weights `[0.25, 0.75]` give one entry `(5, 0.75)`; a sum would give 1.0.
#[test]
fn a_repeated_token_keeps_its_largest_weight() {
  let weight = Tensor::new(&[[1f32]], &Device::Cpu).unwrap();
  let bias = Tensor::new(&[0f32], &Device::Cpu).unwrap();
  let model = head(weight, bias, 64);

  let hidden = Tensor::new(&[[[0.25f32], [0.75]]], &Device::Cpu).unwrap();
  let input_ids = Tensor::from_vec(vec![5u32, 5], (1, 2), &Device::Cpu).unwrap();
  let rows = model.scatter_rows(&hidden, &input_ids, &[2]).unwrap();

  assert_eq!(rows[0].indices, vec![5]);
  assert!(
    (rows[0].values[0] - 0.75).abs() < 1e-6,
    "{:?}",
    rows[0].values
  );
}

/// The reference zeroes `<s>`, `</s>`, `<pad>` and `<unk>`, ids 0, 2, 1 and 3, after the scatter.
#[test]
fn the_four_special_ids_are_dropped() {
  let weight = Tensor::new(&[[1f32]], &Device::Cpu).unwrap();
  let bias = Tensor::new(&[0.1f32], &Device::Cpu).unwrap();
  let model = head(weight, bias, 64);

  let hidden = Tensor::new(
    &[[[0.1f32], [0.2], [0.3], [0.4], [0.5], [0.6]]],
    &Device::Cpu,
  )
  .unwrap();
  let input_ids = Tensor::from_vec(vec![0u32, 2, 1, 3, 9, 42], (1, 6), &Device::Cpu).unwrap();
  let rows = model.scatter_rows(&hidden, &input_ids, &[6]).unwrap();

  assert_eq!(rows[0].indices, vec![9, 42]);
  for (got, want) in rows[0].values.iter().zip([0.6f32, 0.7]) {
    assert!((got - want).abs() < 1e-6, "{got} vs {want}");
  }
}

/// `bge-m3`'s `empty` fixture tokenizes to `<s> </s>`, whose weights are all 0 after `relu`.
#[test]
fn an_all_zero_row_is_an_empty_embedding_not_an_error() {
  let weight = Tensor::new(&[[0f32]], &Device::Cpu).unwrap();
  let bias = Tensor::new(&[0f32], &Device::Cpu).unwrap();
  let model = head(weight, bias, 64);

  let hidden = Tensor::new(&[[[1f32], [2.0]]], &Device::Cpu).unwrap();
  let input_ids = Tensor::from_vec(vec![9u32, 10], (1, 2), &Device::Cpu).unwrap();
  let rows = model.scatter_rows(&hidden, &input_ids, &[2]).unwrap();

  assert!(rows[0].is_empty(), "{:?}", rows[0]);
}

/// The padded positions carry id 999 and a large weight, which must not reach the output.
#[test]
fn padded_positions_never_reach_the_fold() {
  let weight = Tensor::new(&[[1f32]], &Device::Cpu).unwrap();
  let bias = Tensor::new(&[0f32], &Device::Cpu).unwrap();
  let model = head(weight, bias, 64);

  let hidden = Tensor::new(&[[[0.1f32], [0.5], [0.05], [9.0], [9.0]]], &Device::Cpu).unwrap();
  let input_ids = Tensor::from_vec(vec![0u32, 55, 2, 999, 999], (1, 5), &Device::Cpu).unwrap();
  let rows = model.scatter_rows(&hidden, &input_ids, &[3]).unwrap();

  assert_eq!(rows[0].indices, vec![55]);
  assert!(
    (rows[0].values[0] - 0.5).abs() < 1e-6,
    "{:?}",
    rows[0].values
  );
}

/// A `(1, 256, vocab)` tensor fits in memory, so only the output shape shows no tensor takes `vocab_size` as a dimension.
#[test]
fn vocab_size_never_sizes_a_tensor() {
  let weight = Tensor::new(&[[1f32]], &Device::Cpu).unwrap();
  let bias = Tensor::new(&[0f32], &Device::Cpu).unwrap();
  let model = head(weight, bias, 250_002);

  let seq = 256;
  let hidden = Tensor::from_vec(vec![0.5f32; seq], (1, seq, 1), &Device::Cpu).unwrap();
  let ids: Vec<u32> = (0..seq as u32).collect();
  let input_ids = Tensor::from_vec(ids, (1, seq), &Device::Cpu).unwrap();
  let rows = model.scatter_rows(&hidden, &input_ids, &[seq]).unwrap();

  assert_eq!(rows[0].len(), seq - 4);
}

/// FlagEmbedding's own loader hides a missing file behind an info log and a randomly initialized head (`runner.py:90-97`).
#[test]
fn a_missing_sparse_linear_file_is_a_named_load_error() {
  let dir = tempdir().unwrap();
  let missing = dir.path().join("sparse_linear.pt");
  let varmap = VarMap::new();
  let vb = VarBuilder::from_varmap(&varmap, DType::F32, &Device::Cpu);

  let err = M3Sparse::load(xlm_config(tiny_config()), vb, &missing).unwrap_err();
  assert!(
    matches!(err, CanddingError::MissingFile(ref p) if p == &missing),
    "{err}"
  );
  assert!(err.to_string().contains("sparse_linear.pt"), "{err}");
}

/// `linear_head.bin` matches `sparse_linear.pt`'s one-column shape, so it serves this loader as it serves `M3Colbert`'s.
#[test]
fn load_reads_weight_and_bias_from_a_real_archive() {
  let fixture = std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
    .join("tests")
    .join("fixtures")
    .join("pickle")
    .join("linear_head.bin");
  let config = xlm_config(BertConfig {
    hidden_size: 4,
    ..tiny_config()
  });
  let varmap = VarMap::new();
  let vb = VarBuilder::from_varmap(&varmap, DType::F32, &Device::Cpu);
  let model = M3Sparse::load(config, vb, &fixture).unwrap();

  assert_eq!(
    model
      .weight
      .to_dtype(DType::F32)
      .unwrap()
      .to_vec2::<f32>()
      .unwrap(),
    vec![vec![1.5f32, -2.0, 0.5, 0.25]]
  );
  assert_eq!(
    model
      .bias
      .to_dtype(DType::F32)
      .unwrap()
      .to_vec1::<f32>()
      .unwrap(),
    vec![0.75f32]
  );

  let input_ids = Tensor::from_vec(vec![4u32, 5, 6, 7], (1, 4), &Device::Cpu).unwrap();
  let attention_mask = Tensor::from_vec(vec![1u32; 4], (1, 4), &Device::Cpu).unwrap();
  let token_type_ids = input_ids.zeros_like().unwrap();
  let input = EncoderInput {
    input_ids: &input_ids,
    attention_mask: &attention_mask,
    token_type_ids: &token_type_ids,
    lengths: &[4],
  };
  model.forward(&input).unwrap();
}

/// The weights start at f16, as the pickle reader hands them; skipping the cast would fail an f32 by f16 matmul.
#[test]
fn the_head_is_cast_from_f16_to_the_runs_dtype() {
  let weight = Tensor::new(&[[0.5f32, -0.25, 0.1, 0.2]], &Device::Cpu)
    .unwrap()
    .to_dtype(DType::F16)
    .unwrap();
  let bias = Tensor::new(&[0.1f32], &Device::Cpu)
    .unwrap()
    .to_dtype(DType::F16)
    .unwrap();
  assert_eq!(weight.dtype(), DType::F16);

  let config = xlm_config(BertConfig {
    hidden_size: 4,
    ..tiny_config()
  });
  let varmap = VarMap::new();
  let vb = VarBuilder::from_varmap(&varmap, DType::F32, &Device::Cpu);
  let model = M3Sparse::from_raw_head(config, vb, weight, bias).unwrap();
  assert_eq!(model.weight.dtype(), DType::F32);
  assert_eq!(model.bias.dtype(), DType::F32);

  let input_ids = Tensor::from_vec(vec![4u32, 5, 6, 7], (1, 4), &Device::Cpu).unwrap();
  let attention_mask = Tensor::from_vec(vec![1u32; 4], (1, 4), &Device::Cpu).unwrap();
  let token_type_ids = input_ids.zeros_like().unwrap();
  let input = EncoderInput {
    input_ids: &input_ids,
    attention_mask: &attention_mask,
    token_type_ids: &token_type_ids,
    lengths: &[4],
  };
  model.forward(&input).unwrap();
}