realizar 0.8.5

Pure Rust ML inference engine built from scratch - model serving for GGUF and safetensors
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
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266


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

    #[test]
    fn test_config_to_gpu() {
        let apr_config = AprTransformerConfig {
            architecture: "test".to_string(),
            hidden_dim: 512,
            num_layers: 4,
            num_heads: 8,
            num_kv_heads: 4,
            vocab_size: 32000,
            intermediate_dim: 1024,
            context_length: 2048,
            rope_theta: 10000.0,
            eps: 1e-5,
            eos_token_id: None,
        ..Default::default()
        };

        let gpu_config = AprToGpuAdapter::config_to_gpu(&apr_config);

        assert_eq!(gpu_config.vocab_size, 32000);
        assert_eq!(gpu_config.hidden_dim, 512);
        assert_eq!(gpu_config.num_heads, 8);
        assert_eq!(gpu_config.num_kv_heads, 4);
        assert_eq!(gpu_config.num_layers, 4);
        assert_eq!(gpu_config.intermediate_dim, 1024);
        assert_eq!(gpu_config.eps, 1e-5);
    }

    #[test]
    fn test_transpose_matrix() {
        let data = vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]; // 2x3
        let transposed = transpose_matrix(&data, 2, 3); // 3x2

        assert_eq!(transposed, vec![1.0, 4.0, 2.0, 5.0, 3.0, 6.0]);
    }

    #[test]
    fn test_transpose_identity() {
        let data = vec![1.0, 2.0, 3.0, 4.0]; // 2x2
        let transposed = transpose_matrix(&data, 2, 2);

        assert_eq!(transposed, vec![1.0, 3.0, 2.0, 4.0]);
    }

    #[test]
    fn test_transpose_single_row() {
        let data = vec![1.0, 2.0, 3.0, 4.0]; // 1x4
        let transposed = transpose_matrix(&data, 1, 4); // 4x1

        assert_eq!(transposed, vec![1.0, 2.0, 3.0, 4.0]);
    }

    #[test]
    fn test_transpose_single_col() {
        let data = vec![1.0, 2.0, 3.0, 4.0]; // 4x1
        let transposed = transpose_matrix(&data, 4, 1); // 1x4

        assert_eq!(transposed, vec![1.0, 2.0, 3.0, 4.0]);
    }

    #[test]
    fn test_transpose_3x4() {
        let data: Vec<f32> = (1..=12).map(|x| x as f32).collect(); // 3x4
        let transposed = transpose_matrix(&data, 3, 4); // 4x3

        // Original: [[1,2,3,4], [5,6,7,8], [9,10,11,12]]
        // Transposed: [[1,5,9], [2,6,10], [3,7,11], [4,8,12]]
        assert_eq!(transposed[0], 1.0);
        assert_eq!(transposed[1], 5.0);
        assert_eq!(transposed[2], 9.0);
        assert_eq!(transposed[3], 2.0);
        assert_eq!(transposed[4], 6.0);
    }

    #[test]
    fn test_transpose_double_transpose_is_identity() {
        let data: Vec<f32> = (1..=20).map(|x| x as f32).collect();
        let rows = 4;
        let cols = 5;

        let transposed1 = transpose_matrix(&data, rows, cols);
        let transposed2 = transpose_matrix(&transposed1, cols, rows);

        assert_eq!(transposed2, data);
    }

    #[test]
    fn test_apr_gpu_error_dequant_display() {
        let err = AprGpuError::DequantError("test error".to_string());
        let display = format!("{}", err);
        assert!(display.contains("dequantize"));
        assert!(display.contains("test error"));
    }

    #[test]
    fn test_apr_gpu_error_dimension_mismatch_display() {
        let err = AprGpuError::DimensionMismatch {
            expected: 100,
            actual: 50,
        };
        let display = format!("{}", err);
        assert!(display.contains("dimension mismatch"));
        assert!(display.contains("100"));
        assert!(display.contains("50"));
    }

    #[test]
    fn test_apr_gpu_error_gpu_model_error_display() {
        let err = AprGpuError::GpuModelError("creation failed".to_string());
        let display = format!("{}", err);
        assert!(display.contains("GpuModel"));
        assert!(display.contains("creation failed"));
    }

    #[test]
    fn test_config_to_gpu_with_gqa() {
        // Test Grouped Query Attention config
        let apr_config = AprTransformerConfig {
            architecture: "llama".to_string(),
            hidden_dim: 4096,
            num_layers: 32,
            num_heads: 32,
            num_kv_heads: 8, // GQA: fewer KV heads
            vocab_size: 32000,
            intermediate_dim: 11008,
            context_length: 4096,
            rope_theta: 10000.0,
            eps: 1e-5,
            eos_token_id: None,
        ..Default::default()
        };

        let gpu_config = AprToGpuAdapter::config_to_gpu(&apr_config);

        assert_eq!(gpu_config.num_heads, 32);
        assert_eq!(gpu_config.num_kv_heads, 8);
        // Verify GQA ratio
        assert_eq!(gpu_config.num_heads / gpu_config.num_kv_heads, 4);
    }

    #[test]
    fn test_config_to_gpu_with_mha() {
        // Test Multi-Head Attention config (num_heads == num_kv_heads)
        let apr_config = AprTransformerConfig {
            architecture: "phi2".to_string(),
            hidden_dim: 2560,
            num_layers: 32,
            num_heads: 32,
            num_kv_heads: 32, // MHA: same as num_heads
            vocab_size: 51200,
            intermediate_dim: 10240,
            context_length: 2048,
            rope_theta: 10000.0,
            eps: 1e-5,
            eos_token_id: None,
        ..Default::default()
        };

        let gpu_config = AprToGpuAdapter::config_to_gpu(&apr_config);

        assert_eq!(gpu_config.num_heads, gpu_config.num_kv_heads);
    }

    #[test]
    fn test_apr_gpu_error_debug() {
        let err = AprGpuError::DequantError("debug test".to_string());
        let debug = format!("{:?}", err);
        assert!(debug.contains("DequantError"));
        assert!(debug.contains("debug test"));
    }

    #[test]
    fn test_transpose_preserves_sum() {
        let data: Vec<f32> = (1..=24).map(|x| x as f32).collect();
        let transposed = transpose_matrix(&data, 4, 6);

        let original_sum: f32 = data.iter().sum();
        let transposed_sum: f32 = transposed.iter().sum();
        assert!((original_sum - transposed_sum).abs() < 1e-6);
    }

    #[test]
    fn test_config_to_gpu_hybrid_fields_passthrough() {
        // GH-278: Verify hybrid attention fields propagate through config_to_gpu
        let apr_config = AprTransformerConfig {
            architecture: "qwen3_5".to_string(),
            hidden_dim: 5120,
            num_layers: 64,
            num_heads: 40,
            num_kv_heads: 8,
            vocab_size: 248064,
            intermediate_dim: 25600,
            context_length: 131072,
            rope_theta: 1_300_000.0,
            eps: 1e-6,
            eos_token_id: Some(248044),
            explicit_head_dim: Some(256),
            layer_types: Some(vec![
                "full_attention".to_string(),
                "linear_attention".to_string(),
            ]),
            linear_key_head_dim: Some(128),
            linear_value_head_dim: Some(128),
            linear_num_key_heads: Some(16),
            linear_num_value_heads: Some(48),
            linear_conv_kernel_dim: Some(4),
            num_experts: None,
            num_experts_per_tok: None,
            expert_intermediate_size: None,
        };

        let gpu_config = AprToGpuAdapter::config_to_gpu(&apr_config);

        assert_eq!(gpu_config.explicit_head_dim, Some(256));
        assert_eq!(gpu_config.head_dim(), 256);
        assert_eq!(gpu_config.linear_key_head_dim, Some(128));
        assert_eq!(gpu_config.linear_value_head_dim, Some(128));
        assert_eq!(gpu_config.linear_num_key_heads, Some(16));
        assert_eq!(gpu_config.linear_num_value_heads, Some(48));
        assert_eq!(gpu_config.linear_conv_kernel_dim, Some(4));
        // Verify layer_types propagated
        let lt = gpu_config.layer_types.as_ref().expect("layer_types");
        assert_eq!(lt.len(), 2);
        assert!(!gpu_config.is_linear_layer(0)); // full_attention
        assert!(gpu_config.is_linear_layer(1)); // linear_attention
        // Verify derived dimensions
        assert_eq!(gpu_config.linear_key_dim(), 16 * 128); // 2048
        assert_eq!(gpu_config.linear_value_dim(), 48 * 128); // 6144
        assert_eq!(gpu_config.linear_conv_dim(), 2 * 2048 + 6144); // 10240
    }

    #[test]
    fn test_config_to_gpu_no_hybrid_fields() {
        // Backwards compat: standard model has all hybrid fields as None
        let apr_config = AprTransformerConfig {
            architecture: "llama".to_string(),
            hidden_dim: 4096,
            num_layers: 32,
            num_heads: 32,
            num_kv_heads: 8,
            vocab_size: 32000,
            intermediate_dim: 11008,
            context_length: 4096,
            rope_theta: 10000.0,
            eps: 1e-5,
            eos_token_id: None,
            ..Default::default()
        };

        let gpu_config = AprToGpuAdapter::config_to_gpu(&apr_config);

        assert!(gpu_config.explicit_head_dim.is_none());
        assert!(gpu_config.layer_types.is_none());
        assert!(gpu_config.linear_key_head_dim.is_none());
        assert!(!gpu_config.is_linear_layer(0));
        assert_eq!(gpu_config.head_dim(), 128); // computed: 4096/32
        assert_eq!(gpu_config.linear_key_dim(), 0);
        assert_eq!(gpu_config.linear_value_dim(), 0);
    }
}