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
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295


// ============================================================================
// Tests (Protocol T-COV-95)
// ============================================================================

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

    // === apply_rope tests ===

    #[test]
    fn test_apply_rope_basic() {
        // Single position, single head, head_dim=2
        let mut x = vec![1.0, 0.0]; // [cos, sin] position encoding input
        let seq_len = 1;
        let num_heads = 1;
        let head_dim = 2;
        let rope_theta = 10000.0;
        let start_pos = 0;

        apply_rope(&mut x, seq_len, num_heads, head_dim, rope_theta, start_pos);

        // At position 0, angle = 0 * freq = 0, so sin(0)=0, cos(0)=1
        // Rotation: [x1*1 - x2*0, x1*0 + x2*1] = [1, 0]
        assert!((x[0] - 1.0).abs() < 1e-5, "x[0] = {}", x[0]);
        assert!(x[1].abs() < 1e-5, "x[1] = {}", x[1]);
    }

    #[test]
    fn test_apply_rope_position_one() {
        // Single position at pos=1 should show rotation
        let mut x = vec![1.0, 0.0];
        apply_rope(&mut x, 1, 1, 2, 10000.0, 1);

        // At position 1, angle = 1 * 1/(10000^0) = 1
        // cos(1) ≈ 0.54, sin(1) ≈ 0.84
        let expected_cos = 1.0f32.cos();
        let expected_sin = 1.0f32.sin();
        assert!(
            (x[0] - expected_cos).abs() < 1e-5,
            "x[0] = {}, expected {}",
            x[0],
            expected_cos
        );
        assert!(
            (x[1] - expected_sin).abs() < 1e-5,
            "x[1] = {}, expected {}",
            x[1],
            expected_sin
        );
    }

    #[test]
    fn test_apply_rope_multiple_positions() {
        // Two positions
        let mut x = vec![
            1.0, 0.0, // Position 0
            1.0, 0.0, // Position 1
        ];
        apply_rope(&mut x, 2, 1, 2, 10000.0, 0);

        // Position 0: no rotation (angle=0)
        assert!((x[0] - 1.0).abs() < 1e-5);
        assert!(x[1].abs() < 1e-5);

        // Position 1: rotated
        let expected_cos = 1.0f32.cos();
        let expected_sin = 1.0f32.sin();
        assert!((x[2] - expected_cos).abs() < 1e-5);
        assert!((x[3] - expected_sin).abs() < 1e-5);
    }

    #[test]
    fn test_apply_rope_multiple_heads() {
        // Single position, two heads
        let mut x = vec![
            1.0, 0.0, // Head 0
            0.0, 1.0, // Head 1
        ];
        apply_rope(&mut x, 1, 2, 2, 10000.0, 0);

        // At position 0, both heads get identity rotation
        assert!((x[0] - 1.0).abs() < 1e-5);
        assert!(x[1].abs() < 1e-5);
        assert!(x[2].abs() < 1e-5);
        assert!((x[3] - 1.0).abs() < 1e-5);
    }

    #[test]
    fn test_apply_rope_larger_head_dim() {
        // Head dim 4 (2 pairs to rotate)
        let mut x = vec![1.0, 2.0, 3.0, 4.0];
        apply_rope(&mut x, 1, 1, 4, 10000.0, 0);

        // At position 0, angle = 0, no rotation
        assert!((x[0] - 1.0).abs() < 1e-5);
        assert!((x[1] - 2.0).abs() < 1e-5);
        assert!((x[2] - 3.0).abs() < 1e-5);
        assert!((x[3] - 4.0).abs() < 1e-5);
    }

    #[test]
    fn test_apply_rope_with_start_pos() {
        let mut x = vec![1.0, 0.0];
        apply_rope(&mut x, 1, 1, 2, 10000.0, 5);

        // At position 5, angle = 5 * 1 = 5
        let expected_cos = 5.0f32.cos();
        let expected_sin = 5.0f32.sin();
        assert!(
            (x[0] - expected_cos).abs() < 1e-5,
            "x[0] = {}, expected {}",
            x[0],
            expected_cos
        );
        assert!(
            (x[1] - expected_sin).abs() < 1e-5,
            "x[1] = {}, expected {}",
            x[1],
            expected_sin
        );
    }

    // === argmax tests ===

    #[test]
    fn test_argmax_kv_single() {
        assert_eq!(argmax(&[5.0]), 0);
    }

    #[test]
    fn test_argmax_kv_first() {
        assert_eq!(argmax(&[10.0, 5.0, 3.0]), 0);
    }

    #[test]
    fn test_argmax_kv_last() {
        assert_eq!(argmax(&[1.0, 2.0, 3.0]), 2);
    }

    #[test]
    fn test_argmax_kv_middle() {
        assert_eq!(argmax(&[1.0, 10.0, 3.0]), 1);
    }

    #[test]
    fn test_argmax_kv_negatives() {
        assert_eq!(argmax(&[-5.0, -2.0, -10.0]), 1);
    }

    #[test]
    fn test_argmax_kv_empty_returns_zero() {
        assert_eq!(argmax(&[]), 0);
    }

    // === sample_topk tests ===

    #[test]
    fn test_sample_topk_returns_max_with_low_temp() {
        let logits = vec![1.0, 10.0, 2.0];
        // With temperature=1.0 and top_k=1, should return argmax
        let result = sample_topk(&logits, 1.0, 1);
        assert_eq!(result, 1);
    }

    #[test]
    fn test_sample_topk_top_3() {
        let logits = vec![0.0, 10.0, 5.0, 1.0];
        // Top-3: indices 1, 2, 3. Should return one of them (likely 1)
        let result = sample_topk(&logits, 1.0, 3);
        assert!(result <= 3, "result = {}", result);
    }

    #[test]
    fn test_sample_topk_with_high_temp() {
        let logits = vec![0.0, 1.0, 0.0];
        // High temperature makes distribution flatter, but max still most likely
        let result = sample_topk(&logits, 10.0, 3);
        assert!(result <= 2);
    }

    #[test]
    fn test_sample_topk_top_1_is_argmax() {
        let logits = vec![0.0, 0.0, 100.0, 0.0];
        let result = sample_topk(&logits, 1.0, 1);
        assert_eq!(result, 2);
    }

    #[test]
    fn test_sample_topk_empty_returns_zero() {
        let result = sample_topk(&[], 1.0, 10);
        assert_eq!(result, 0);
    }

    // === GQA attention dimension tests ===

    #[test]
    fn test_gqa_attention_dimension_calculations() {
        // Test dimension calculations used in GQA attention
        let num_heads = 32;
        let num_kv_heads = 8;
        let head_dim = 128;
        let seq_len = 10;

        let hidden_dim = num_heads * head_dim;
        assert_eq!(hidden_dim, 4096);

        let kv_dim = num_kv_heads * head_dim;
        assert_eq!(kv_dim, 1024);

        let heads_per_kv = num_heads / num_kv_heads;
        assert_eq!(heads_per_kv, 4);

        // Output size
        let output_size = seq_len * hidden_dim;
        assert_eq!(output_size, 40960);

        // Q, K, V sizes
        let q_size = seq_len * hidden_dim;
        let k_size = seq_len * kv_dim;
        let v_size = seq_len * kv_dim;
        assert_eq!(q_size, 40960);
        assert_eq!(k_size, 10240);
        assert_eq!(v_size, 10240);
    }

    #[test]
    fn test_gqa_attention_scale_factor() {
        let head_dim = 128;
        let scale = 1.0 / (head_dim as f32).sqrt();
        let expected = 1.0 / 128.0_f32.sqrt();
        assert!((scale - expected).abs() < 1e-6);
        assert!((scale - 0.088388).abs() < 1e-5);
    }

    // === gqa_incremental_attention tests ===

    #[test]
    fn test_gqa_incremental_attention_params() {
        // Just test the function exists and has correct signature
        // Actually calling it would require a valid GpuModel
        let num_heads = 4;
        let num_kv_heads = 2;
        let head_dim = 16;
        let cache_len = 5;

        // These would be inputs
        let _q = vec![0.0f32; num_heads * head_dim];
        let _all_k = vec![0.0f32; cache_len * num_kv_heads * head_dim];
        let _all_v = vec![0.0f32; cache_len * num_kv_heads * head_dim];

        // Verify dimension calculations
        let hidden_dim = num_heads * head_dim;
        assert_eq!(hidden_dim, 64);
        let kv_dim = num_kv_heads * head_dim;
        assert_eq!(kv_dim, 32);
        let heads_per_kv = num_heads / num_kv_heads;
        assert_eq!(heads_per_kv, 2);
    }

    // === layer_norm_kv smoke test ===

    #[test]
    fn test_layer_norm_static_dimensions() {
        // Test that GpuModel::layer_norm_static returns correct dimensions
        let input = vec![1.0, 2.0, 3.0, 4.0];
        let weight = vec![1.0; 4];
        let bias = vec![0.0; 4];
        let hidden_dim = 4;
        let eps = 1e-5;

        let output = GpuModel::layer_norm_static(&input, &weight, &bias, hidden_dim, eps);
        assert_eq!(output.len(), 4);

        // Verify output is finite
        for &v in &output {
            assert!(v.is_finite(), "output contains non-finite value: {}", v);
        }
    }

    #[test]
    fn test_layer_norm_static_preserves_length() {
        // Test that layer norm preserves the input length
        let hidden_dim = 8;
        let seq_len = 4;
        let input = vec![1.0; seq_len * hidden_dim];
        let weight = vec![1.0; hidden_dim];
        let bias = vec![0.0; hidden_dim];
        let eps = 1e-5;

        let output = GpuModel::layer_norm_static(&input, &weight, &bias, hidden_dim, eps);
        assert_eq!(output.len(), seq_len * hidden_dim);
    }
}