trustformers-models 0.1.1

Model implementations for TrustformeRS
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

use serde::{Deserialize, Serialize};
use trustformers_core::traits::Config;

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct DebertaConfig {
    pub vocab_size: usize,
    pub hidden_size: usize,
    pub num_hidden_layers: usize,
    pub num_attention_heads: usize,
    pub intermediate_size: usize,
    pub hidden_act: String,
    pub hidden_dropout_prob: f32,
    pub attention_probs_dropout_prob: f32,
    pub max_position_embeddings: usize,
    pub type_vocab_size: usize,
    pub initializer_range: f32,
    pub layer_norm_eps: f32,
    pub pad_token_id: u32,
    pub position_embedding_type: String,
    pub use_cache: bool,
    pub classifier_dropout: Option<f32>,

    // DeBERTa-specific parameters
    pub relative_attention: bool,
    pub max_relative_positions: i32,
    pub pos_att_type: Vec<String>, // ["p2c", "c2p"] for positional attention types
    pub norm_rel_ebd: String,      // "layer_norm" or "none"
    pub share_att_key: bool,
    pub model_type: String,
}

impl Default for DebertaConfig {
    fn default() -> Self {
        Self {
            vocab_size: 50265,
            hidden_size: 768,
            num_hidden_layers: 12,
            num_attention_heads: 12,
            intermediate_size: 3072,
            hidden_act: "gelu".to_string(),
            hidden_dropout_prob: 0.1,
            attention_probs_dropout_prob: 0.1,
            max_position_embeddings: 2048,
            type_vocab_size: 0, // DeBERTa doesn't use token type embeddings by default
            initializer_range: 0.02,
            layer_norm_eps: 1e-7,
            pad_token_id: 0,
            position_embedding_type: "relative_key_query".to_string(),
            use_cache: true,
            classifier_dropout: None,

            // DeBERTa-specific defaults
            relative_attention: true,
            max_relative_positions: -1, // -1 means no limit
            pos_att_type: vec!["p2c".to_string(), "c2p".to_string()],
            norm_rel_ebd: "layer_norm".to_string(),
            share_att_key: true,
            model_type: "deberta".to_string(),
        }
    }
}

impl Config for DebertaConfig {
    fn validate(&self) -> trustformers_core::errors::Result<()> {
        if !self.hidden_size.is_multiple_of(self.num_attention_heads) {
            return Err(trustformers_core::errors::invalid_config(
                "hidden_size",
                "hidden_size must be divisible by num_attention_heads",
            ));
        }

        if self.max_relative_positions == 0 {
            return Err(trustformers_core::errors::invalid_config(
                "max_relative_positions",
                "max_relative_positions cannot be 0",
            ));
        }

        Ok(())
    }

    fn architecture(&self) -> &'static str {
        "DeBERTa"
    }
}

impl DebertaConfig {
    /// DeBERTa-Base configuration
    pub fn base() -> Self {
        Self {
            vocab_size: 50265,
            hidden_size: 768,
            num_hidden_layers: 12,
            num_attention_heads: 12,
            intermediate_size: 3072,
            hidden_act: "gelu".to_string(),
            hidden_dropout_prob: 0.1,
            attention_probs_dropout_prob: 0.1,
            max_position_embeddings: 2048,
            type_vocab_size: 0,
            initializer_range: 0.02,
            layer_norm_eps: 1e-7,
            pad_token_id: 0,
            position_embedding_type: "relative_key_query".to_string(),
            use_cache: true,
            classifier_dropout: None,

            relative_attention: true,
            max_relative_positions: -1,
            pos_att_type: vec!["p2c".to_string(), "c2p".to_string()],
            norm_rel_ebd: "layer_norm".to_string(),
            share_att_key: true,
            model_type: "deberta".to_string(),
        }
    }

    /// DeBERTa-Large configuration
    pub fn large() -> Self {
        Self {
            vocab_size: 50265,
            hidden_size: 1024,
            num_hidden_layers: 24,
            num_attention_heads: 16,
            intermediate_size: 4096,
            hidden_act: "gelu".to_string(),
            hidden_dropout_prob: 0.1,
            attention_probs_dropout_prob: 0.1,
            max_position_embeddings: 2048,
            type_vocab_size: 0,
            initializer_range: 0.02,
            layer_norm_eps: 1e-7,
            pad_token_id: 0,
            position_embedding_type: "relative_key_query".to_string(),
            use_cache: true,
            classifier_dropout: None,

            relative_attention: true,
            max_relative_positions: -1,
            pos_att_type: vec!["p2c".to_string(), "c2p".to_string()],
            norm_rel_ebd: "layer_norm".to_string(),
            share_att_key: true,
            model_type: "deberta".to_string(),
        }
    }

    /// DeBERTa-v2-XLarge configuration
    pub fn xlarge() -> Self {
        Self {
            vocab_size: 128100,
            hidden_size: 1536,
            num_hidden_layers: 24,
            num_attention_heads: 24,
            intermediate_size: 6144,
            hidden_act: "gelu".to_string(),
            hidden_dropout_prob: 0.1,
            attention_probs_dropout_prob: 0.1,
            max_position_embeddings: 2048,
            type_vocab_size: 0,
            initializer_range: 0.02,
            layer_norm_eps: 1e-7,
            pad_token_id: 0,
            position_embedding_type: "relative_key_query".to_string(),
            use_cache: true,
            classifier_dropout: None,

            relative_attention: true,
            max_relative_positions: -1,
            pos_att_type: vec!["p2c".to_string(), "c2p".to_string()],
            norm_rel_ebd: "layer_norm".to_string(),
            share_att_key: true,
            model_type: "deberta".to_string(),
        }
    }

    /// DeBERTa-v3-Large configuration
    pub fn v3_large() -> Self {
        Self {
            vocab_size: 128100,
            hidden_size: 1024,
            num_hidden_layers: 24,
            num_attention_heads: 16,
            intermediate_size: 4096,
            hidden_act: "gelu".to_string(),
            hidden_dropout_prob: 0.1,
            attention_probs_dropout_prob: 0.1,
            max_position_embeddings: 2048,
            type_vocab_size: 0,
            initializer_range: 0.02,
            layer_norm_eps: 1e-7,
            pad_token_id: 0,
            position_embedding_type: "relative_key_query".to_string(),
            use_cache: true,
            classifier_dropout: None,

            relative_attention: true,
            max_relative_positions: -1,
            pos_att_type: vec!["p2c".to_string(), "c2p".to_string()],
            norm_rel_ebd: "layer_norm".to_string(),
            share_att_key: true,
            model_type: "deberta-v2".to_string(),
        }
    }

    /// Create DeBERTa configuration from model name
    pub fn from_pretrained_name(model_name: &str) -> Self {
        let name_lower = model_name.to_lowercase();

        if name_lower.contains("xlarge") || name_lower.contains("xxl") {
            Self::xlarge()
        } else if name_lower.contains("large") {
            if name_lower.contains("v3") {
                Self::v3_large()
            } else {
                Self::large()
            }
        } else {
            // Default to base for unknown variants
            Self::base()
        }
    }
}