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

use serde::Deserialize;

use crate::layers::activation::Activation;

#[derive(Clone, Copy, Debug, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case")]
pub enum LayerType {
  SlidingAttention,
  FullAttention,
}

#[derive(Clone, Debug, Deserialize)]
pub struct Gemma3Config {
  pub vocab_size: usize,
  pub hidden_size: usize,
  pub num_hidden_layers: usize,
  pub num_attention_heads: usize,
  pub num_key_value_heads: usize,
  pub head_dim: usize,
  pub intermediate_size: usize,
  #[serde(default = "default_act")]
  pub hidden_activation: Activation,
  pub max_position_embeddings: usize,
  #[serde(default = "default_eps")]
  pub rms_norm_eps: f64,
  #[serde(default = "default_global_theta")]
  pub rope_theta: f64,
  #[serde(default = "default_local_theta")]
  pub rope_local_base_freq: f64,
  #[serde(default = "default_window")]
  pub sliding_window: usize,
  #[serde(default)]
  pub layer_types: Option<Vec<LayerType>>,
  #[serde(default, rename = "_sliding_window_pattern")]
  pub sliding_window_pattern: Option<usize>,
  pub query_pre_attn_scalar: f64,
  #[serde(default)]
  pub use_bidirectional_attention: bool,
  #[serde(default)]
  pub attention_bias: bool,
  #[serde(default)]
  pub attn_logit_softcapping: Option<f64>,
  #[serde(default)]
  pub rope_scaling: Option<serde_json::Value>,
}

fn default_act() -> Activation {
  Activation::GeluPytorchTanh
}

fn default_eps() -> f64 {
  1e-6
}

fn default_global_theta() -> f64 {
  1_000_000.0
}

fn default_local_theta() -> f64 {
  10_000.0
}

fn default_window() -> usize {
  512
}

impl Gemma3Config {
  /// `layer_types` from the config, else every `pattern`-th layer (default 6) is global, the rest sliding.
  pub fn layer_types(&self) -> Vec<LayerType> {
    if let Some(types) = &self.layer_types {
      return types.clone();
    }
    let pattern = self.sliding_window_pattern.unwrap_or(6).max(1);
    (0..self.num_hidden_layers)
      .map(|i| {
        if (i + 1) % pattern == 0 {
          LayerType::FullAttention
        } else {
          LayerType::SlidingAttention
        }
      })
      .collect()
  }

  /// transformers halves the stored span and adds one when bidirectional: a stored 512 becomes `|i - j| < 257`.
  pub fn attention_window(&self) -> usize {
    if self.use_bidirectional_attention {
      self.sliding_window / 2 + 1
    } else {
      self.sliding_window
    }
  }
}

#[cfg(test)]
mod tests {
  use super::*;

  fn config(sliding_window: usize) -> Gemma3Config {
    Gemma3Config {
      vocab_size: 1,
      hidden_size: 1,
      num_hidden_layers: 1,
      num_attention_heads: 1,
      num_key_value_heads: 1,
      head_dim: 1,
      intermediate_size: 1,
      hidden_activation: Activation::GeluPytorchTanh,
      max_position_embeddings: 1,
      rms_norm_eps: 1e-6,
      rope_theta: 1.0,
      rope_local_base_freq: 1.0,
      sliding_window,
      layer_types: None,
      sliding_window_pattern: None,
      query_pre_attn_scalar: 1.0,
      use_bidirectional_attention: true,
      attention_bias: false,
      attn_logit_softcapping: None,
      rope_scaling: None,
    }
  }

  #[test]
  fn attention_window_halves_the_stored_span_and_adds_one() {
    assert_eq!(config(512).attention_window(), 257);
    assert_eq!(config(6).attention_window(), 4);
    assert_eq!(config(513).attention_window(), 257);
  }

  #[test]
  fn attention_window_keeps_the_stored_span_without_bidirectional_attention() {
    let mut causal = config(512);
    causal.use_bidirectional_attention = false;
    assert_eq!(causal.attention_window(), 512);
  }
}