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

//! `tiny_tokenizer_qwen3_reranker.json`, written and checked by `scripts/prompt_fixture.py`, encodes every text below to the published ids.

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

use candle_core::Device;
use tempfile::TempDir;
use tempfile::tempdir;

use crate::error::CanddingError;
use crate::error::LengthKind;
use crate::tokenizer::PadToken;
use crate::tokenizer::PromptTemplate;
use crate::tokenizer::PromptTokenizer;

pub(super) const PAD: u32 = 151_643;
pub(super) const DEFAULT: &str =
  "Given a web search query, retrieve relevant passages that answer the query";
const CUSTOM: &str =
  "Given a question about geography, judge whether the passage names the correct city";
const QUERY: &str = "What is the capital of China?";
const DOCUMENT: &str = "The capital of China is Beijing.";

/// Identical on sentence-transformers' `CrossEncoder` and on the model card's own code.
const P0_IDS: [u32; 85] = [
  151644, 8948, 198, 60256, 3425, 279, 11789, 20027, 279, 8502, 3118, 389, 279, 11361, 323, 279,
  758, 1235, 3897, 13, 7036, 429, 279, 4226, 646, 1172, 387, 330, 9693, 1, 476, 330, 2152, 3263,
  151645, 198, 151644, 872, 198, 27, 641, 1235, 26818, 16246, 264, 3482, 2711, 3239, 11, 17179,
  9760, 46769, 429, 4226, 279, 3239, 198, 27, 2859, 26818, 3555, 374, 279, 6722, 315, 5616, 5267,
  75692, 26818, 576, 6722, 315, 5616, 374, 26549, 13, 151645, 198, 151644, 77091, 198, 151667, 271,
  151668, 271,
];

pub(super) fn subset_path() -> PathBuf {
  PathBuf::from(env!("CARGO_MANIFEST_DIR"))
    .join("tests/fixtures/tiny_tokenizer_qwen3_reranker.json")
}

pub(super) fn qwen3(path: &Path, max_length: usize) -> PromptTokenizer {
  PromptTokenizer::from_file(
    path,
    max_length,
    PadToken::Id(PAD),
    PromptTemplate::Qwen3Reranker,
  )
  .unwrap()
}

fn edited(edit: impl FnOnce(&mut serde_json::Value)) -> (TempDir, PathBuf) {
  let mut json: serde_json::Value =
    serde_json::from_str(&fs::read_to_string(subset_path()).unwrap()).unwrap();
  edit(&mut json);
  let dir = tempdir().unwrap();
  let path = dir.path().join("tokenizer.json");
  fs::write(&path, serde_json::to_string(&json).unwrap()).unwrap();
  (dir, path)
}

#[test]
fn the_template_renders_its_prefix_content_and_suffix_verbatim() {
  let template = PromptTemplate::Qwen3Reranker;
  assert_eq!(
    template.render("I", "Q", "D"),
    format!(
      "{}<Instruct>: I\n<Query>: Q\n<Document>: D{}",
      template.prefix(),
      template.suffix()
    )
  );
  assert!(template.suffix().ends_with("</think>\n\n"));
  assert!(!template.suffix().ends_with("\n\n\n"));
}

/// The known triple encodes to the references' 85 ids: 39 of prefix, the content and 9 of suffix.
#[test]
fn a_known_triple_encodes_to_the_references_ids() {
  let tokenizer = qwen3(&subset_path(), 8192);
  assert_eq!(tokenizer.prefix_len(), 39);
  assert_eq!(tokenizer.suffix_ids(), &P0_IDS[76..]);
  assert_eq!(
    tokenizer.prompt_ids(DEFAULT, QUERY, DOCUMENT).unwrap(),
    P0_IDS
  );
}

#[test]
fn the_instruction_reaches_its_slot() {
  let tokenizer = qwen3(&subset_path(), 8192);
  let slot = |instruction: &str| {
    let ids = tokenizer.prompt_ids(instruction, QUERY, DOCUMENT).unwrap();
    let head = &P0_IDS[..43];
    let tail = &P0_IDS[56..];
    assert!(ids.starts_with(head), "{instruction:?}");
    assert!(ids.ends_with(tail), "{instruction:?}");
    ids[head.len()..ids.len() - tail.len()].to_vec()
  };
  assert_eq!(slot(DEFAULT), &P0_IDS[43..56]);
  let custom = slot(CUSTOM);
  assert_ne!(custom, slot(DEFAULT));
  assert_eq!(custom.len(), 14);
}

/// The space before the newline encodes with it: one id where the default instruction took 13.
#[test]
fn an_empty_instruction_renders_empty() {
  let template = PromptTemplate::Qwen3Reranker;
  assert!(
    template
      .content("", QUERY, DOCUMENT)
      .starts_with("<Instruct>: \n<Query>: ")
  );
  let tokenizer = qwen3(&subset_path(), 8192);
  let empty = tokenizer.prompt_ids("", QUERY, DOCUMENT).unwrap();
  assert_eq!(empty.len(), P0_IDS.len() - 13);
  assert_eq!(&empty[..43], &P0_IDS[..43]);
  assert_eq!(&empty[44..], &P0_IDS[57..]);
}

#[test]
fn a_max_length_below_the_templates_minimum_is_refused_by_name() {
  for max_length in [0, 1, 9, 39, 48] {
    let refused = PromptTokenizer::from_file(
      &subset_path(),
      max_length,
      PadToken::Id(PAD),
      PromptTemplate::Qwen3Reranker,
    );
    match refused {
      Err(
        err @ CanddingError::LengthTooSmall {
          requested,
          minimum: 49,
          kind:
            LengthKind::Prompt {
              template: PromptTemplate::Qwen3Reranker,
              prefix: 39,
              suffix: 9,
            },
        },
      ) => {
        assert_eq!(requested, max_length);
        let message = err.to_string();
        assert!(message.contains("minimum 49"), "{message}");
        assert!(message.contains("qwen3_reranker"), "{message}");
      }
      Err(other) => panic!("{max_length}: unexpected error: {other}"),
      Ok(_) => panic!("{max_length} leaves no content token"),
    }
  }
  assert_eq!(qwen3(&subset_path(), 49).max_length(), 49);
}

/// A prompt rendered with the template file's `\n\n\n` suffix ends in that one id instead of the suffix's `\n\n`.
#[test]
fn the_suffix_check_refuses_a_rendering_that_disagrees_with_the_template() {
  let tokenizer = qwen3(&subset_path(), 8192);
  let template = PromptTemplate::Qwen3Reranker;
  let rendered = format!("{}\n", template.render(DEFAULT, QUERY, DOCUMENT));
  match tokenizer.rendered_ids(&rendered) {
    Err(CanddingError::PromptSuffixMismatch {
      expected, found, ..
    }) => {
      assert_eq!(expected, &P0_IDS[76..]);
      assert_ne!(found, expected);
    }
    other => panic!("expected a suffix mismatch, got {other:?}"),
  }
}

/// Without the special tokens the merges join the document's final `.` with `<`, and as text they push the minimum above 49.
#[test]
fn the_suffix_check_refuses_a_tokenizer_that_merges_the_suffix_into_the_document() {
  let (_dir, path) = edited(|json| json["added_tokens"] = serde_json::json!([]));
  let tokenizer = qwen3(&path, 8192);
  let minimum = tokenizer.prefix_len() + tokenizer.suffix_ids().len() + 1;
  assert!(minimum > 49, "{minimum}");
  let refused =
    PromptTokenizer::from_file(&path, 49, PadToken::Id(PAD), PromptTemplate::Qwen3Reranker);
  assert!(
    matches!(refused, Err(CanddingError::LengthTooSmall { minimum: m, kind: LengthKind::Prompt { .. }, .. }) if m == minimum),
    "{:?}",
    refused.err()
  );
  match tokenizer.prompt_ids(DEFAULT, QUERY, DOCUMENT) {
    Err(CanddingError::PromptSuffixMismatch {
      expected, found, ..
    }) => assert_ne!(found, expected),
    other => panic!("expected a suffix mismatch, got {other:?}"),
  }
}

/// The published file's no-op post-processor is not what keeps added tokens out of a prompt.
#[test]
fn a_prompt_gets_no_special_tokens_from_the_post_processor() {
  let (_dir, path) = edited(|json| {
    json["post_processor"] = serde_json::json!({
      "type": "TemplateProcessing",
      "single": [{"SpecialToken": {"id": "<|endoftext|>", "type_id": 0}}, {"Sequence": {"id": "A", "type_id": 0}}],
      "pair": [{"Sequence": {"id": "A", "type_id": 0}}, {"Sequence": {"id": "B", "type_id": 1}}],
      "special_tokens": {"<|endoftext|>": {"id": "<|endoftext|>", "ids": [PAD], "tokens": ["<|endoftext|>"]}}
    })
  });
  let tokenizer = qwen3(&path, 8192);
  assert_eq!(
    tokenizer.prompt_ids(DEFAULT, QUERY, DOCUMENT).unwrap(),
    P0_IDS
  );
}

#[test]
fn a_batch_is_right_padded_with_real_lengths_and_zero_token_types() {
  let tokenizer = qwen3(&subset_path(), 8192);
  let pairs = [(QUERY, DOCUMENT), ("", "")];
  let batch = tokenizer.encode(DEFAULT, &pairs, &Device::Cpu).unwrap();
  let short = tokenizer.prompt_ids(DEFAULT, "", "").unwrap();
  assert_eq!(batch.lengths, vec![85, short.len()]);
  assert_eq!(tokenizer.count(DEFAULT, &pairs).unwrap(), batch.lengths);
  let ids = batch.input_ids.to_vec2::<u32>().unwrap();
  assert_eq!(ids[0], P0_IDS);
  assert_eq!(&ids[1][..short.len()], short.as_slice());
  assert!(ids[1][short.len()..].iter().all(|&id| id == PAD));
  let mask = batch.attention_mask.to_vec2::<u32>().unwrap();
  assert_eq!(mask[1].iter().filter(|&&m| m == 1).count(), short.len());
  assert!(mask[0].iter().all(|&m| m == 1));
  let types = batch.token_type_ids.to_vec2::<u32>().unwrap();
  assert!(types.iter().flatten().all(|&t| t == 0));
}