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


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

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

    // =========================================================================
    // Attention Tests
    // =========================================================================

    #[test]
    fn test_attention_new() {
        let attn = Attention::new(64).expect("valid head_dim");
        assert_eq!(attn.head_dim, 64);
        // Scale should be 1/sqrt(64) = 1/8 = 0.125
        assert!((attn.scale - 0.125).abs() < 1e-6);
    }

    #[test]
    fn test_attention_new_zero_head_dim() {
        let result = Attention::new(0);
        assert!(result.is_err());
    }

    #[test]
    fn test_attention_scale_calculation() {
        // head_dim=16: scale = 1/sqrt(16) = 1/4 = 0.25
        let attn = Attention::new(16).expect("valid");
        assert!((attn.scale - 0.25).abs() < 1e-6);

        // head_dim=64: scale = 1/sqrt(64) = 1/8 = 0.125
        let attn = Attention::new(64).expect("valid");
        assert!((attn.scale - 0.125).abs() < 1e-6);

        // head_dim=128: scale = 1/sqrt(128) ≈ 0.0884
        let attn = Attention::new(128).expect("valid");
        let expected = 1.0 / (128.0_f32).sqrt();
        assert!((attn.scale - expected).abs() < 1e-6);
    }

    #[test]
    fn test_attention_forward_simple() {
        let attn = Attention::new(4).expect("valid");

        // Create simple 2x4 tensors (seq_len=2, head_dim=4)
        let q = Tensor::from_vec(vec![2, 4], vec![1.0; 8]).expect("valid");
        let k = Tensor::from_vec(vec![2, 4], vec![1.0; 8]).expect("valid");
        let v = Tensor::from_vec(vec![2, 4], vec![1.0; 8]).expect("valid");

        let result = attn.forward(&q, &k, &v);
        assert!(result.is_ok());

        let output = result.expect("should succeed");
        assert_eq!(output.shape(), &[2, 4]);
    }

    #[test]
    fn test_attention_forward_shape_mismatch() {
        let attn = Attention::new(4).expect("valid");

        let q = Tensor::from_vec(vec![2, 4], vec![1.0; 8]).expect("valid");
        let k = Tensor::from_vec(vec![2, 8], vec![1.0; 16]).expect("valid"); // Wrong dim
        let v = Tensor::from_vec(vec![2, 4], vec![1.0; 8]).expect("valid");

        let result = attn.forward(&q, &k, &v);
        assert!(result.is_err());
    }

    #[test]
    fn test_attention_forward_1d_tensor() {
        let attn = Attention::new(4).expect("valid");

        // 1D tensor works (treated as seq_len=1)
        let q = Tensor::from_vec(vec![4], vec![1.0; 4]).expect("valid");
        let k = Tensor::from_vec(vec![4], vec![1.0; 4]).expect("valid");
        let v = Tensor::from_vec(vec![4], vec![1.0; 4]).expect("valid");

        let result = attn.forward(&q, &k, &v);
        assert!(result.is_ok());
    }

    // =========================================================================
    // SlidingWindowAttention Tests
    // =========================================================================

    #[test]
    fn test_sliding_window_attention_new() {
        let attn = SlidingWindowAttention::new(64, 512).expect("valid");
        assert_eq!(attn.head_dim(), 64);
        assert_eq!(attn.window_size(), 512);
    }

    #[test]
    fn test_sliding_window_attention_zero_head_dim() {
        let result = SlidingWindowAttention::new(0, 512);
        assert!(result.is_err());
    }

    #[test]
    fn test_sliding_window_attention_zero_window() {
        let result = SlidingWindowAttention::new(64, 0);
        assert!(result.is_err());
    }

    #[test]
    fn test_sliding_window_attention_forward() {
        let attn = SlidingWindowAttention::new(4, 8).expect("valid");

        let q = Tensor::from_vec(vec![2, 4], vec![1.0; 8]).expect("valid");
        let k = Tensor::from_vec(vec![2, 4], vec![1.0; 8]).expect("valid");
        let v = Tensor::from_vec(vec![2, 4], vec![1.0; 8]).expect("valid");

        let result = attn.forward(&q, &k, &v);
        assert!(result.is_ok());

        let output = result.expect("should succeed");
        assert_eq!(output.shape(), &[2, 4]);
    }

    // =========================================================================
    // FusedQKVAttention Tests
    // =========================================================================

    #[test]
    fn test_fused_qkv_attention_new() {
        // FusedQKVAttention::new(head_dim, hidden_dim)
        // hidden_dim must be divisible by head_dim
        let attn = FusedQKVAttention::new(64, 256).expect("valid");
        assert_eq!(attn.head_dim(), 64);
    }

    #[test]
    fn test_fused_qkv_attention_zero_head_dim() {
        let result = FusedQKVAttention::new(0, 256);
        assert!(result.is_err());
    }

    #[test]
    fn test_fused_qkv_attention_zero_hidden_dim() {
        let result = FusedQKVAttention::new(64, 0);
        assert!(result.is_err());
    }

    #[test]
    fn test_fused_qkv_attention_hidden_not_divisible() {
        // hidden_dim=100 is not divisible by head_dim=64
        let result = FusedQKVAttention::new(64, 100);
        assert!(result.is_err());
    }

    #[test]
    fn test_fused_qkv_attention_forward() {
        // FusedQKVAttention takes input [seq_len, hidden_dim] and projects Q, K, V internally
        let attn = FusedQKVAttention::new(4, 8).expect("valid"); // head_dim=4, hidden_dim=8

        // Input: [seq_len=2, hidden_dim=8]
        let input = Tensor::from_vec(vec![2, 8], vec![1.0; 16]).expect("valid");

        let result = attn.forward(&input);
        assert!(result.is_ok());

        let output = result.expect("should succeed");
        assert_eq!(output.shape(), &[2, 8]); // Output: [seq_len, hidden_dim]
    }

    // =========================================================================
    // MultiHeadAttention Tests
    // =========================================================================

    #[test]
    fn test_mha_new() {
        // MultiHeadAttention::new(hidden_dim, num_heads, num_kv_heads)
        // head_dim = hidden_dim / num_heads = 256 / 4 = 64
        let mha = MultiHeadAttention::new(256, 4, 4).expect("valid");
        assert_eq!(mha.num_heads(), 4);
        assert_eq!(mha.num_kv_heads(), 4);
        assert_eq!(mha.head_dim(), 64);
        assert_eq!(mha.hidden_dim(), 256);
    }

    #[test]
    fn test_mha_gqa() {
        // GQA: 8 query heads, 2 KV heads
        // head_dim = 512 / 8 = 64
        let mha = MultiHeadAttention::new(512, 8, 2).expect("valid");
        assert!(mha.is_gqa());
        assert!(!mha.is_mqa());
        assert!(!mha.is_mha());
    }

    #[test]
    fn test_mha_mqa() {
        // MQA: 8 query heads, 1 KV head
        let mha = MultiHeadAttention::new(512, 8, 1).expect("valid");
        assert!(mha.is_mqa());
        assert!(!mha.is_gqa());
        assert!(!mha.is_mha());
    }

    #[test]
    fn test_mha_standard() {
        // Standard MHA: num_heads == num_kv_heads
        let mha = MultiHeadAttention::new(512, 8, 8).expect("valid");
        assert!(mha.is_mha());
        assert!(!mha.is_gqa());
        assert!(!mha.is_mqa());
    }

    #[test]
    fn test_mha_zero_hidden_dim() {
        let result = MultiHeadAttention::new(0, 4, 4);
        assert!(result.is_err());
    }

    #[test]
    fn test_mha_zero_heads() {
        let result = MultiHeadAttention::new(256, 0, 4);
        assert!(result.is_err());
    }

    #[test]
    fn test_mha_zero_kv_heads() {
        let result = MultiHeadAttention::new(256, 4, 0);
        assert!(result.is_err());
    }

    #[test]
    fn test_mha_hidden_dim_not_divisible() {
        // hidden_dim=101 is not divisible by num_heads=4
        let result = MultiHeadAttention::new(101, 4, 4);
        assert!(result.is_err());
    }

    #[test]
    fn test_mha_kv_heads_not_divisor() {
        // num_heads=8 must be divisible by num_kv_heads=3, but 8 % 3 != 0
        let result = MultiHeadAttention::new(512, 8, 3);
        assert!(result.is_err());
    }

    #[test]
    fn test_mha_kv_heads_greater_than_heads() {
        // num_kv_heads=16 > num_heads=8 is invalid
        let result = MultiHeadAttention::new(512, 8, 16);
        assert!(result.is_err());
    }
}