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

use std::fs;
use std::path::PathBuf;

use tempfile::tempdir;

use super::*;
use crate::error::LengthKind;

mod budgets;
mod markers;
mod pairs;
mod prompt;
mod prompt_truncation;
mod windows;

fn tiny(max_length: usize) -> TextTokenizer {
  let path = PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("tests/fixtures/tiny_tokenizer.json");
  TextTokenizer::from_file(&path, max_length, PadToken::Id(0)).unwrap()
}

#[test]
fn pads_on_the_right_and_reports_lengths() {
  let batch = tiny(64)
    .encode(&["hello world", "the cat sat on the mat"], &Device::Cpu)
    .unwrap();
  assert_eq!(batch.input_ids.dims(), &[2, 8]);
  let ids = batch.input_ids.to_vec2::<u32>().unwrap();
  assert_eq!(ids[0], vec![2, 4, 5, 3, 0, 0, 0, 0]);
  assert_eq!(ids[1], vec![2, 6, 7, 8, 9, 6, 10, 3]);
  let mask = batch.attention_mask.to_vec2::<u32>().unwrap();
  assert_eq!(mask[0], vec![1, 1, 1, 1, 0, 0, 0, 0]);
  assert_eq!(
    batch.token_type_ids.to_vec2::<u32>().unwrap()[1],
    vec![0; 8]
  );
  assert_eq!(batch.lengths, vec![4, 8]);
}

#[test]
fn truncates_to_max_length_including_special_tokens() {
  let tokenizer = tiny(5);
  let batch = tokenizer
    .encode(&["hello world", "the cat sat on the mat"], &Device::Cpu)
    .unwrap();
  assert_eq!(batch.input_ids.dims(), &[2, 5]);
  assert_eq!(
    batch.input_ids.to_vec2::<u32>().unwrap()[1],
    vec![2, 6, 7, 8, 3]
  );
  assert_eq!(batch.lengths, vec![4, 5]);
  assert_eq!(
    tokenizer.count(&["the cat sat on the mat", ""]).unwrap(),
    vec![5, 2]
  );
}

/// Counting must not pad: one long document beside thousands of short texts would materialize them all at its length.
#[test]
fn counting_encodes_without_padding() {
  let tokenizer = tiny(64);
  assert!(
    tokenizer.inner.get_padding().is_none(),
    "the tokenizer that counts must carry no padding"
  );
  let texts = ["hello", "the cat sat on the mat", "hello world"];
  assert_eq!(tokenizer.count(&texts).unwrap(), vec![3, 8, 4]);
  let padded = tokenizer.encode(&texts, &Device::Cpu).unwrap();
  assert_eq!(padded.input_ids.dims(), &[3, 8], "encode still pads");
  assert_eq!(padded.lengths, vec![3, 8, 4]);
}

/// The tokenizers crate subtracts the special tokens from `max_length` unchecked, panicking in debug and wrapping in release.
#[test]
fn a_max_length_that_leaves_no_room_for_text_is_refused() {
  let path = PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("tests/fixtures/tiny_tokenizer.json");
  for max_length in [0, 1, 2] {
    let err = match TextTokenizer::from_file(&path, max_length, PadToken::Id(0)) {
      Ok(_) => panic!("max_length {max_length} must be refused"),
      Err(err) => err,
    };
    assert!(
      matches!(err, CanddingError::LengthTooSmall { requested, minimum: 3, kind: LengthKind::Text { added: 2 } }
        if requested == max_length),
      "{max_length}: {err}"
    );
    assert!(err.to_string().contains('3'), "{err}");
  }
  assert_eq!(tiny(3).count(&["hello world"]).unwrap(), vec![3]);
}

#[test]
fn splits_subwords() {
  let ids = tiny(16)
    .encode(&["cats"], &Device::Cpu)
    .unwrap()
    .input_ids
    .to_vec2::<u32>()
    .unwrap();
  assert_eq!(ids[0], vec![2, 7, 11, 3]);
}

#[test]
fn named_pad_token_is_looked_up_in_the_vocabulary() {
  let path = PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("tests/fixtures/tiny_tokenizer.json");
  let tokenizer =
    TextTokenizer::from_file(&path, 64, PadToken::Named("[SEP]".to_string())).unwrap();
  let ids = tokenizer
    .encode(&["hello", "the cat sat"], &Device::Cpu)
    .unwrap()
    .input_ids
    .to_vec2::<u32>()
    .unwrap();
  assert_eq!(ids[0], vec![2, 4, 3, 3, 3]);
}

#[test]
fn pad_token_prefers_config_then_tokenizer_config_then_unknown() {
  let dir = tempdir().unwrap();
  let plain = dir.path().join("plain.json");
  fs::write(&plain, r#"{"pad_token": "<|endoftext|>"}"#).unwrap();
  let added = dir.path().join("added.json");
  fs::write(
    &added,
    r#"{"pad_token": {"content": "<pad>", "lstrip": false}}"#,
  )
  .unwrap();
  assert!(matches!(
    read_pad_token(r#"{"pad_token_id": 1}"#, Some(plain.as_path())).unwrap(),
    PadToken::Id(1)
  ));
  assert!(matches!(
    read_pad_token(r#"{"pad_token_id": null}"#, Some(plain.as_path())).unwrap(),
    PadToken::Named(name) if name == "<|endoftext|>"
  ));
  assert!(matches!(
    read_pad_token("{}", Some(added.as_path())).unwrap(),
    PadToken::Named(name) if name == "<pad>"
  ));
  assert!(matches!(
    read_pad_token("{}", None).unwrap(),
    PadToken::Unknown
  ));
}

/// `"pad_token_id": -1` means none and is not a `u32`, so the priority order carries on to `tokenizer_config.json`.
#[test]
fn a_negative_pad_token_id_is_absent_rather_than_fatal() {
  let dir = tempdir().unwrap();
  let named = dir.path().join("tokenizer_config.json");
  fs::write(&named, r#"{"pad_token": "<pad>"}"#).unwrap();
  assert!(matches!(
    read_pad_token(r#"{"pad_token_id": -1}"#, Some(named.as_path())).unwrap(),
    PadToken::Named(name) if name == "<pad>"
  ));
  assert!(matches!(
    read_pad_token(r#"{"pad_token_id": -1}"#, None).unwrap(),
    PadToken::Unknown
  ));
  assert!(matches!(
    read_pad_token(r#"{"pad_token_id": 0}"#, Some(named.as_path())).unwrap(),
    PadToken::Id(0),
  ));
}

#[test]
fn read_mask_token_prefers_the_published_name_and_falls_back_to_mask_bracket() {
  let dir = tempdir().unwrap();
  let named = dir.path().join("tokenizer_config.json");
  fs::write(&named, r#"{"mask_token": "<mask>"}"#).unwrap();
  assert_eq!(read_mask_token(Some(named.as_path())).unwrap(), "<mask>");

  let added = dir.path().join("added.json");
  fs::write(
    &added,
    r#"{"mask_token": {"content": "<mask>", "lstrip": true}}"#,
  )
  .unwrap();
  assert_eq!(read_mask_token(Some(added.as_path())).unwrap(), "<mask>");

  let no_mask_key = dir.path().join("no_mask_key.json");
  fs::write(&no_mask_key, r#"{"pad_token": "[PAD]"}"#).unwrap();
  assert_eq!(
    read_mask_token(Some(no_mask_key.as_path())).unwrap(),
    "[MASK]"
  );

  assert_eq!(read_mask_token(None).unwrap(), "[MASK]");
}

/// `Tokenizer::from_file` installs a published `padding` section, which the counting tokenizer must clear.
#[test]
fn a_padding_section_in_the_published_file_is_not_inherited() {
  let source = PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("tests/fixtures/tiny_tokenizer.json");
  let mut json: serde_json::Value =
    serde_json::from_str(&fs::read_to_string(source).unwrap()).unwrap();
  json["padding"] = serde_json::json!({
    "strategy": {"Fixed": 8},
    "direction": "Right",
    "pad_to_multiple_of": null,
    "pad_id": 0,
    "pad_type_id": 0,
    "pad_token": "[PAD]"
  });
  let dir = tempfile::tempdir().unwrap();
  let path = dir.path().join("tokenizer.json");
  fs::write(&path, json.to_string()).unwrap();
  let tokenizer = TextTokenizer::from_file(&path, 64, PadToken::Id(0)).unwrap();
  assert!(
    tokenizer.inner.get_padding().is_none(),
    "the file's padding must be cleared"
  );
  assert_eq!(tokenizer.count(&["", "hello"]).unwrap(), vec![2, 3]);
  let batch = tokenizer.encode(&["hello"], &Device::Cpu).unwrap();
  assert_eq!(
    batch.input_ids.dims(),
    &[1, 3],
    "a batch pads to its longest row, not to the file's fixed length"
  );
}