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

use std::collections::HashMap;

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

use super::*;
use crate::error::CanddingError;

mod alibi;
mod flash;

/// Its sigmoid, `0.880797`, is far from the raw `2.0` an `Identity` head returns.
const BIAS: f32 = 2.0;
const SIGMOID_OF_BIAS: f32 = 0.880_797_1;

fn zeros(shape: &[usize]) -> Tensor {
  Tensor::zeros(shape, DType::F32, &Device::Cpu).unwrap()
}

/// A zero-layer encoder's embeddings under `prefix`, every tensor zero except the LayerNorm weight.
fn embeddings(prefix: &str, type_vocab_size: usize) -> HashMap<String, Tensor> {
  let mut tensors = HashMap::new();
  for (name, shape) in [
    ("embeddings.word_embeddings.weight", vec![8, 4]),
    ("embeddings.position_embeddings.weight", vec![16, 4]),
    (
      "embeddings.token_type_embeddings.weight",
      vec![type_vocab_size, 4],
    ),
    ("embeddings.LayerNorm.bias", vec![4]),
  ] {
    tensors.insert(format!("{prefix}.{name}"), zeros(&shape));
  }
  tensors.insert(
    format!("{prefix}.embeddings.LayerNorm.weight"),
    Tensor::ones(4, DType::F32, &Device::Cpu).unwrap(),
  );
  tensors
}

fn bias() -> Tensor {
  Tensor::new(&[BIAS], &Device::Cpu).unwrap()
}

fn score(model: &dyn RerankModel) -> Vec<f32> {
  let input_ids = Tensor::new(&[[1u32, 2, 3, 4], [1u32, 5, 3, 0]], &Device::Cpu).unwrap();
  let attention_mask = Tensor::new(&[[1u32, 1, 1, 1], [1u32, 1, 1, 0]], &Device::Cpu).unwrap();
  let token_type_ids = input_ids.zeros_like().unwrap();
  let lengths = vec![4, 3];
  let input = EncoderInput {
    input_ids: &input_ids,
    attention_mask: &attention_mask,
    token_type_ids: &token_type_ids,
    lengths: &lengths,
  };
  model
    .forward(&input)
    .unwrap()
    .iter()
    .map(RerankScore::value)
    .collect()
}

fn conventions(head: RerankHead, score_activation: ScoreActivation) -> RerankConventions {
  RerankConventions {
    head,
    score_activation,
    alibi_variant: None,
    position_scheme: None,
  }
}

#[test]
fn the_bert_pooler_head_applies_the_activation_load_rerank_hands_it() {
  let mut tensors = embeddings("bert", 2);
  tensors.insert("bert.pooler.dense.weight".to_string(), zeros(&[4, 4]));
  tensors.insert("bert.pooler.dense.bias".to_string(), zeros(&[4]));
  tensors.insert("classifier.weight".to_string(), zeros(&[1, 4]));
  tensors.insert("classifier.bias".to_string(), bias());
  let vb = VarBuilder::from_tensors(tensors, DType::F32, &Device::Cpu);
  let config_json = r#"{"vocab_size":8,"hidden_size":4,"num_hidden_layers":0,
    "num_attention_heads":1,"intermediate_size":1,"max_position_embeddings":16,
    "type_vocab_size":2,"layer_norm_eps":1e-12}"#;
  let model = load_rerank(
    "bert",
    config_json,
    vb,
    conventions(RerankHead::BertPooler, ScoreActivation::Sigmoid),
  )
  .unwrap();
  for score in score(model.as_ref()) {
    assert!((score - SIGMOID_OF_BIAS).abs() < 1e-6, "{score}");
  }
}

#[test]
fn the_roberta_classification_head_applies_the_activation_load_rerank_hands_it() {
  let mut tensors = embeddings("roberta", 1);
  tensors.insert("classifier.dense.weight".to_string(), zeros(&[4, 4]));
  tensors.insert("classifier.dense.bias".to_string(), zeros(&[4]));
  tensors.insert("classifier.out_proj.weight".to_string(), zeros(&[1, 4]));
  tensors.insert("classifier.out_proj.bias".to_string(), bias());
  let vb = VarBuilder::from_tensors(tensors, DType::F32, &Device::Cpu);
  let config_json = r#"{"vocab_size":8,"hidden_size":4,"num_hidden_layers":0,
    "num_attention_heads":1,"intermediate_size":1,"max_position_embeddings":16,
    "type_vocab_size":1,"layer_norm_eps":1e-12,"pad_token_id":1,
    "position_embedding_type":"absolute"}"#;
  let model = load_rerank(
    "xlm-roberta",
    config_json,
    vb,
    conventions(
      RerankHead::RobertaClassificationHead,
      ScoreActivation::Sigmoid,
    ),
  )
  .unwrap();
  for score in score(model.as_ref()) {
    assert!((score - SIGMOID_OF_BIAS).abs() < 1e-6, "{score}");
  }
}

#[test]
fn head_serializes_with_a_kind_tag() {
  assert_eq!(
    serde_json::to_value(RerankHead::RobertaClassificationHead).unwrap(),
    serde_json::json!({"kind": "roberta_classification_head"})
  );
  for (unembedding, name) in [(Unembedding::Tied, "tied"), (Unembedding::Untied, "untied")] {
    let logit_score = RerankHead::LogitScore {
      true_token_id: 9693,
      false_token_id: 2152,
      unembedding,
    };
    let json = serde_json::json!(
      {"kind": "logit_score", "true_token_id": 9693, "false_token_id": 2152, "unembedding": name}
    );
    assert_eq!(serde_json::to_value(logit_score).unwrap(), json);
    assert_eq!(
      serde_json::from_value::<RerankHead>(json).unwrap(),
      logit_score
    );
    assert_eq!(unembedding.name(), name);
  }
  assert_eq!(
    serde_json::from_str::<RerankHead>(r#"{"kind": "bert_pooler"}"#).unwrap(),
    RerankHead::BertPooler
  );
}

/// The "yes" row is every token's embedding and the "no" row zero, so a row scores 4, its normalized embedding's squared norm.
#[test]
fn the_logit_score_head_takes_the_ids_and_the_activation_load_rerank_hands_it() {
  let config_json = r#"{"vocab_size":8,"hidden_size":4,"num_hidden_layers":0,"num_attention_heads":1,
    "num_key_value_heads":1,"head_dim":4,"intermediate_size":4,"max_position_embeddings":16,
    "tie_word_embeddings":true}"#;
  let build = |true_token_id: u32, false_token_id: u32, unembedding, activation| {
    let mut embed = vec![1f32; 8 * 4];
    embed[4..8].fill(0.0);
    let mut tensors = HashMap::new();
    tensors.insert(
      "model.embed_tokens.weight".to_string(),
      Tensor::from_vec(embed, (8, 4), &Device::Cpu).unwrap(),
    );
    tensors.insert(
      "model.norm.weight".to_string(),
      Tensor::ones(4, DType::F32, &Device::Cpu).unwrap(),
    );
    let vb = VarBuilder::from_tensors(tensors, DType::F32, &Device::Cpu);
    let head = RerankHead::LogitScore {
      true_token_id,
      false_token_id,
      unembedding,
    };
    load_rerank("qwen3", config_json, vb, conventions(head, activation))
  };
  let score_with = |true_token_id: u32, false_token_id: u32, activation: ScoreActivation| {
    let model = build(true_token_id, false_token_id, Unembedding::Tied, activation).unwrap();
    score(model.as_ref())
  };
  let untied = build(3, 1, Unembedding::Untied, ScoreActivation::Identity);
  assert!(
    matches!(
      untied,
      Err(CanddingError::UnembeddingDisagreesWithConfig { .. })
    ),
    "{:?}",
    untied.err()
  );
  for value in score_with(3, 1, ScoreActivation::Identity) {
    assert!((value - 4.0).abs() < 1e-4, "{value}");
  }
  for value in score_with(1, 3, ScoreActivation::Identity) {
    assert!((value + 4.0).abs() < 1e-4, "{value}");
  }
  for value in score_with(3, 1, ScoreActivation::Sigmoid) {
    assert!(
      (value - 1.0 / (1.0 + (-4.0f32).exp())).abs() < 1e-5,
      "{value}"
    );
  }
}