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

//! `naver/splade-code-06B`'s tokenizer adds no special tokens, so the empty string has zero positions.

use super::*;

fn empty_hidden(batch: usize, hidden_size: usize) -> Tensor {
  Tensor::from_vec(Vec::<f32>::new(), (batch, 0, hidden_size), &Device::Cpu).unwrap()
}

fn empty_mask(batch: usize) -> Tensor {
  Tensor::from_vec(Vec::<u32>::new(), (batch, 0), &Device::Cpu).unwrap()
}

#[test]
fn pool_logits_returns_empty_rows_for_a_zero_length_sequence() {
  let model = tiny_model();
  let rows = model
    .pool_logits(&empty_hidden(1, 8), &empty_mask(1))
    .unwrap();
  assert_eq!(rows.len(), 1);
  assert!(rows[0].is_empty(), "{:?}", rows[0]);
}

#[test]
fn pool_logits_handles_every_row_of_a_multi_row_zero_length_batch() {
  let model = tiny_model();
  let rows = model
    .pool_logits(&empty_hidden(3, 8), &empty_mask(3))
    .unwrap();
  assert_eq!(rows.len(), 3);
  assert!(rows.iter().all(SparseEmbedding::is_empty));
}

#[test]
fn forward_returns_an_empty_row_for_a_zero_token_batch_of_one() {
  let model = tiny_model();
  let input_ids = Tensor::from_vec(Vec::<u32>::new(), (1, 0), &Device::Cpu).unwrap();
  let attention_mask = Tensor::from_vec(Vec::<u32>::new(), (1, 0), &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: &[0],
  };
  let rows = model.forward(&input).unwrap();
  assert_eq!(rows.len(), 1);
  assert!(rows[0].is_empty());
}

#[test]
fn a_zero_length_row_among_real_rows_matches_the_row_encoded_alone() {
  let model = tiny_model();
  let device = Device::Cpu;

  let alone_ids = Tensor::from_vec(vec![1u32, 2], (1, 2), &device).unwrap();
  let alone_mask = Tensor::from_vec(vec![1u32, 1], (1, 2), &device).unwrap();
  let alone_types = alone_ids.zeros_like().unwrap();
  let alone = model
    .forward(&EncoderInput {
      input_ids: &alone_ids,
      attention_mask: &alone_mask,
      token_type_ids: &alone_types,
      lengths: &[2],
    })
    .unwrap();

  let input_ids = Tensor::from_vec(vec![1u32, 2, 3, 0], (2, 2), &device).unwrap();
  let attention_mask = Tensor::from_vec(vec![1u32, 1, 0, 0], (2, 2), &device).unwrap();
  let token_type_ids = input_ids.zeros_like().unwrap();
  let rows = model
    .forward(&EncoderInput {
      input_ids: &input_ids,
      attention_mask: &attention_mask,
      token_type_ids: &token_type_ids,
      lengths: &[2, 0],
    })
    .unwrap();

  assert_eq!(rows.len(), 2);
  assert_eq!(
    rows[0].indices, alone[0].indices,
    "{:?} vs {:?}",
    rows[0], alone[0]
  );
  for (a, b) in rows[0].values.iter().zip(&alone[0].values) {
    assert!((a - b).abs() < 1e-5, "{:?} vs {:?}", rows[0], alone[0]);
  }
  assert!(rows[1].is_empty(), "{:?}", rows[1]);
}

#[test]
fn real_rows_on_either_side_of_an_empty_one_each_match_their_own_alone_encoding() {
  let model = tiny_model();
  let device = Device::Cpu;

  let alone_row = |ids: &[u32]| {
    let input_ids = Tensor::from_vec(ids.to_vec(), (1, ids.len()), &device).unwrap();
    let attention_mask = Tensor::from_vec(vec![1u32; ids.len()], (1, ids.len()), &device).unwrap();
    let token_type_ids = input_ids.zeros_like().unwrap();
    model
      .forward(&EncoderInput {
        input_ids: &input_ids,
        attention_mask: &attention_mask,
        token_type_ids: &token_type_ids,
        lengths: &[ids.len()],
      })
      .unwrap()
      .remove(0)
  };
  let alone_first = alone_row(&[1, 2]);
  let alone_third = alone_row(&[4, 5, 3]);

  // Row 1 (the empty one) is padded to the batch's longest real row, 3.
  let input_ids = Tensor::from_vec(vec![1u32, 2, 0, 0, 0, 0, 4, 5, 3], (3, 3), &device).unwrap();
  let attention_mask =
    Tensor::from_vec(vec![1u32, 1, 0, 0, 0, 0, 1, 1, 1], (3, 3), &device).unwrap();
  let token_type_ids = input_ids.zeros_like().unwrap();
  let rows = model
    .forward(&EncoderInput {
      input_ids: &input_ids,
      attention_mask: &attention_mask,
      token_type_ids: &token_type_ids,
      lengths: &[2, 0, 3],
    })
    .unwrap();

  assert_eq!(rows.len(), 3);
  assert_eq!(rows[0].indices, alone_first.indices, "{:?}", rows[0]);
  for (a, b) in rows[0].values.iter().zip(&alone_first.values) {
    assert!(
      (a - b).abs() < 1e-5,
      "row 0: {:?} vs {:?}",
      rows[0],
      alone_first
    );
  }
  assert!(rows[1].is_empty(), "{:?}", rows[1]);
  assert_eq!(rows[2].indices, alone_third.indices, "{:?}", rows[2]);
  for (a, b) in rows[2].values.iter().zip(&alone_third.values) {
    assert!(
      (a - b).abs() < 1e-5,
      "row 2: {:?} vs {:?}",
      rows[2],
      alone_third
    );
  }
}

/// candle's Metal SDPA divides a zero softmax sum for a row with no visible key, which a zero-length row is.
#[cfg(feature = "metal")]
#[test]
fn a_zero_length_row_among_real_rows_does_not_corrupt_the_real_row_on_metal() {
  let device = candle_core::Device::new_metal(0).expect("a metal device");
  let varmap = VarMap::new();
  let vb = VarBuilder::from_varmap(&varmap, DType::F32, &device);
  let model = SpladeTiedSparse::load(tiny_qwen3_config(), vb).unwrap();

  let alone_ids = Tensor::from_vec(vec![1u32, 2], (1, 2), &device).unwrap();
  let alone_mask = Tensor::from_vec(vec![1u32, 1], (1, 2), &device).unwrap();
  let alone_types = alone_ids.zeros_like().unwrap();
  let alone = model
    .forward(&EncoderInput {
      input_ids: &alone_ids,
      attention_mask: &alone_mask,
      token_type_ids: &alone_types,
      lengths: &[2],
    })
    .unwrap();

  let input_ids = Tensor::from_vec(vec![1u32, 2, 3, 0], (2, 2), &device).unwrap();
  let attention_mask = Tensor::from_vec(vec![1u32, 1, 0, 0], (2, 2), &device).unwrap();
  let token_type_ids = input_ids.zeros_like().unwrap();
  let rows = model
    .forward(&EncoderInput {
      input_ids: &input_ids,
      attention_mask: &attention_mask,
      token_type_ids: &token_type_ids,
      lengths: &[2, 0],
    })
    .unwrap();

  assert_eq!(rows.len(), 2);
  assert!(
    rows[0].values.iter().all(|v| v.is_finite()),
    "real row must stay finite on Metal: {:?}",
    rows[0]
  );
  assert_eq!(
    rows[0].indices, alone[0].indices,
    "{:?} vs {:?}",
    rows[0], alone[0]
  );
  for (a, b) in rows[0].values.iter().zip(&alone[0].values) {
    assert!((a - b).abs() < 1e-4, "{:?} vs {:?}", rows[0], alone[0]);
  }
  assert!(rows[1].is_empty(), "{:?}", rows[1]);
}