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


#[test]
fn test_tiled_single_head_attention_various_tile_sizes() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let seq_len = 4;
    let head_dim = 4;
    let scale = 1.0 / (head_dim as f32).sqrt();

    let q: Vec<f32> = (0..seq_len * head_dim)
        .map(|i| (i % 7) as f32 * 0.1)
        .collect();
    let k: Vec<f32> = (0..seq_len * head_dim)
        .map(|i| (i % 5) as f32 * 0.1)
        .collect();
    let v: Vec<f32> = (0..seq_len * head_dim)
        .map(|i| (i % 3) as f32 * 0.1)
        .collect();

    let standard = model
        .standard_single_head_attention(&q, &k, &v, seq_len, head_dim, scale)
        .expect("expected value");

    for tile_size in [1, 2, 3, 4, 8] {
        let tiled = model
            .tiled_single_head_attention(&q, &k, &v, seq_len, head_dim, scale, tile_size)
            .expect("expected value");
        for (s, t) in standard.iter().zip(tiled.iter()) {
            assert!((s - t).abs() < 1e-4, "tile_size={}", tile_size);
        }
    }
}

// ============================================================================
// tiled_causal_attention tests
// ============================================================================

#[test]
fn test_tiled_causal_attention_basic() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let seq_len = 3;
    let head_dim = 4;
    let scale = 1.0 / (head_dim as f32).sqrt();

    let q = vec![1.0; seq_len * head_dim];
    let k = vec![1.0; seq_len * head_dim];
    let v = vec![1.0; seq_len * head_dim];

    let result = model.tiled_causal_attention(&q, &k, &v, seq_len, head_dim, scale, 2);
    assert!(result.is_ok());
    let output = result.expect("output");
    assert_eq!(output.len(), seq_len * head_dim);
}

#[test]
fn test_tiled_causal_attention_first_position() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    // First position can only attend to itself
    let seq_len = 3;
    let head_dim = 2;
    let scale = 1.0 / (head_dim as f32).sqrt();

    let q = vec![1.0; seq_len * head_dim];
    let k = vec![1.0; seq_len * head_dim];
    // Different V values for each position
    let v = vec![1.0, 1.0, 2.0, 2.0, 3.0, 3.0];

    let result = model.tiled_causal_attention(&q, &k, &v, seq_len, head_dim, scale, 1);
    assert!(result.is_ok());
    let output = result.expect("output");
    // First position output should be first V (since it only attends to itself)
    assert!((output[0] - 1.0).abs() < 1e-5);
    assert!((output[1] - 1.0).abs() < 1e-5);
}

#[test]
fn test_tiled_causal_attention_various_tile_sizes() {
    let config = test_config();
    let model = create_test_model_with_config(&config);

    let seq_len = 4;
    let head_dim = 4;
    let scale = 1.0 / (head_dim as f32).sqrt();

    let q: Vec<f32> = (0..seq_len * head_dim)
        .map(|i| (i % 7) as f32 * 0.1)
        .collect();
    let k: Vec<f32> = (0..seq_len * head_dim)
        .map(|i| (i % 5) as f32 * 0.1)
        .collect();
    let v: Vec<f32> = (0..seq_len * head_dim)
        .map(|i| (i % 3) as f32 * 0.1)
        .collect();

    // Use tile_size=1 as reference
    let reference = model
        .tiled_causal_attention(&q, &k, &v, seq_len, head_dim, scale, 1)
        .expect("expected value");

    for tile_size in [2, 3, 4, 8] {
        let tiled = model
            .tiled_causal_attention(&q, &k, &v, seq_len, head_dim, scale, tile_size)
            .expect("expected value");
        for (r, t) in reference.iter().zip(tiled.iter()) {
            assert!((r - t).abs() < 1e-4, "tile_size={}", tile_size);
        }
    }
}

// ============================================================================
// GPU feature tests (conditional compilation)
// ============================================================================

#[cfg(feature = "gpu")]
mod gpu_tests {
    use super::*;
    use crate::gguf::DispatchMetrics;
    use std::sync::Arc;

    #[test]
    fn test_forward_batch_with_cache_empty_tokens_error() {
        let config = test_config();
        let model = create_test_model_with_config(&config);
        let mut cache = OwnedQuantizedKVCache::from_config(&config, 100);
        let metrics = Arc::new(DispatchMetrics::new());

        let result = model.forward_batch_with_cache(&[], &mut cache, &metrics);
        assert!(result.is_err());
        let err = result.unwrap_err();
        assert!(format!("{:?}", err).contains("empty"));
    }

    #[test]
    fn test_forward_batch_with_cache_single_token() {
        let config = test_config();
        let model = create_test_model_with_config(&config);
        let mut cache = OwnedQuantizedKVCache::from_config(&config, 100);
        let metrics = Arc::new(DispatchMetrics::new());

        let result = model.forward_batch_with_cache(&[1], &mut cache, &metrics);
        assert!(result.is_ok());
        let logits = result.expect("logits");
        assert_eq!(logits.len(), config.vocab_size);
        // Check that CPU dispatch was recorded
        assert!(metrics.cpu_dispatches() > 0);
    }

    #[test]
    fn test_generate_with_batched_prefill_empty_prompt_error() {
        let config = test_config();
        let model = create_test_model_with_config(&config);
        let gen_config = QuantizedGenerateConfig::deterministic(5);
        let metrics = Arc::new(DispatchMetrics::new());

        let result = model.generate_with_batched_prefill(&[], &gen_config, &metrics);
        assert!(result.is_err());
        let err = result.unwrap_err();
        assert!(format!("{:?}", err).contains("empty"));
    }

    #[test]
    fn test_generate_with_batched_prefill_basic() {
        let config = test_config();
        let model = create_test_model_with_config(&config);
        let gen_config = QuantizedGenerateConfig::deterministic(2);
        let metrics = Arc::new(DispatchMetrics::new());

        let result = model.generate_with_batched_prefill(&[1], &gen_config, &metrics);
        assert!(result.is_ok());
        let tokens = result.expect("tokens");
        assert!(!tokens.is_empty());
        assert_eq!(tokens[0], 1);
    }

    #[test]
    fn test_generate_with_batched_prefill_temperature() {
        let config = test_config();
        let model = create_test_model_with_config(&config);
        let gen_config = QuantizedGenerateConfig::default()
            .with_max_tokens(2)
            .with_temperature(0.8)
            .with_top_k(5);
        let metrics = Arc::new(DispatchMetrics::new());

        let result = model.generate_with_batched_prefill(&[1], &gen_config, &metrics);
        assert!(result.is_ok());
    }

    #[test]
    fn test_reshape_for_parallel_heads_basic() {
        let config = test_config();
        let model = create_test_model_with_config(&config);

        let seq_len = 2;
        let num_heads = 4;
        let head_dim = 16;
        let input: Vec<f32> = (0..seq_len * num_heads * head_dim)
            .map(|i| i as f32)
            .collect();

        let result = model.reshape_for_parallel_heads(&input, seq_len, num_heads, head_dim);
        assert!(result.is_ok());
        let reshaped = result.expect("reshaped");
        assert_eq!(reshaped.len(), num_heads * seq_len * head_dim);
    }

    #[test]
    fn test_reshape_for_parallel_heads_invalid_size() {
        let config = test_config();
        let model = create_test_model_with_config(&config);

        let seq_len = 2;
        let num_heads = 4;
        let head_dim = 16;
        // Wrong input size
        let input: Vec<f32> = vec![1.0; 10];

        let result = model.reshape_for_parallel_heads(&input, seq_len, num_heads, head_dim);
        assert!(result.is_err());
    }

    #[test]
    fn test_apply_causal_mask_softmax_basic() {
        let config = test_config();
        let model = create_test_model_with_config(&config);

        let seq_len = 3;
        // 3x3 scores
        let scores: Vec<f32> = vec![1.0, 2.0, 3.0, 1.0, 2.0, 3.0, 1.0, 2.0, 3.0];

        let weights = model.apply_causal_mask_softmax(&scores, seq_len);
        assert_eq!(weights.len(), seq_len * seq_len);

        // First row: only position 0 is unmasked
        assert!((weights[0] - 1.0).abs() < 1e-6);
        assert!((weights[1]).abs() < 1e-6);
        assert!((weights[2]).abs() < 1e-6);

        // Second row: positions 0,1 are unmasked
        let row1_sum = weights[3] + weights[4];
        assert!((row1_sum - 1.0).abs() < 1e-6);
        assert!((weights[5]).abs() < 1e-6);

        // Third row: all positions are unmasked
        let row2_sum = weights[6] + weights[7] + weights[8];
        assert!((row2_sum - 1.0).abs() < 1e-6);
    }
}