aprender-serve 0.64.0

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


/// IMP-101c: Attention with cache produces normalized output
#[test]
fn test_imp_101c_attention_with_cache_softmax_normalized() {
    let config = GGUFConfig {
        architecture: "test".to_string(),
        constraints: crate::gguf::ArchConstraints::from_architecture("test"),
        hidden_dim: 4,
        intermediate_dim: 16,
        num_layers: 1,
        num_heads: 1,
        num_kv_heads: 1,
        vocab_size: 100,
        context_length: 2048,
        eps: 1e-5,
        rope_type: 0,
        rope_theta: 10000.0,
            explicit_head_dim: None,
            query_pre_attn_scalar: None,
        bos_token_id: None,
            eos_token_id: None,
    };

    let model = OwnedQuantizedModel {
        config,
        token_embedding: vec![],
        position_embedding: None,
        layers: vec![],
        encoder_layers: vec![],
        encoder_output_norm_weight: None,
        encoder_output_norm_bias: None,
        output_norm_weight: vec![],
        output_norm_bias: None,
        lm_head_weight: OwnedQuantizedTensor {
            data: vec![],
            in_dim: 4,
            out_dim: 100,
            qtype: 0,
        },
        lm_head_bias: None,
        #[cfg(feature = "cuda")]
        cuda_executor: None,
        #[cfg(feature = "cuda")]
        cuda_kernel_count: std::sync::atomic::AtomicU64::new(0),
        #[cfg(feature = "cuda")]
        cached_weight_names: std::sync::Mutex::new(std::collections::HashSet::new()),
    };

    // Test attention with cache
    // Q = [1, 0, 0, 0], cached K/V for one position, current K/V
    let q = vec![1.0, 0.0, 0.0, 0.0];
    let k_cache = vec![1.0, 0.0, 0.0, 0.0]; // cached position 0
    let v_cache = vec![1.0, 0.0, 0.0, 0.0];
    let current_k = vec![1.0, 0.0, 0.0, 0.0]; // current position 1
    let current_v = vec![0.0, 1.0, 0.0, 0.0];

    let output = model.attention_with_cache(&q, &k_cache, &v_cache, &current_k, &current_v);

    // Output should be weighted combination of v_cache and current_v
    // Both K vectors are identical to Q, so scores are equal -> 50/50 weights
    // Output should be approximately [0.5, 0.5, 0, 0]
    assert_eq!(
        output.len(),
        4,
        "IMP-101c: Output should have hidden_dim elements"
    );

    let sum: f32 = output.iter().sum();
    assert!(
        (sum - 1.0).abs() < 0.1,
        "IMP-101c: Attention output should be normalized weighted sum"
    );
}

/// IMP-101c: Cache handles multiple positions correctly
#[test]
fn test_imp_101c_kv_cache_multiple_positions() {
    let mut cache = OwnedQuantizedKVCache::new(1, 4, 100);

    // Add 3 positions
    for i in 0..3 {
        let k = vec![i as f32; 4];
        let v = vec![(i as f32) * 0.1; 4];
        cache.append(0, &k, &v);
        cache.advance();
    }

    assert_eq!(cache.len(), 3, "IMP-101c: Cache should have 3 positions");

    let k_data = cache.get_k(0);
    assert_eq!(
        k_data.len(),
        12,
        "IMP-101c: K cache should have 3 * 4 = 12 elements"
    );

    // Verify first position K values
    assert!(
        (k_data[0] - 0.0).abs() < 1e-6,
        "IMP-101c: First K should be 0"
    );
    // Verify second position K values
    assert!(
        (k_data[4] - 1.0).abs() < 1e-6,
        "IMP-101c: Second K should be 1"
    );
    // Verify third position K values
    assert!(
        (k_data[8] - 2.0).abs() < 1e-6,
        "IMP-101c: Third K should be 2"
    );
}

#[test]
fn test_imp_105_gqa_attention_multiple_q_per_kv() {
    // IMP-105: GQA (Grouped Query Attention) support
    // 8 Q heads share 2 KV heads (4 Q heads per KV head)
    let config = GGUFConfig {
        architecture: "test".to_string(),
        constraints: crate::gguf::ArchConstraints::from_architecture("test"),
        hidden_dim: 32, // 8 heads * 4 head_dim
        intermediate_dim: 128,
        num_layers: 1,
        num_heads: 8,    // Q heads
        num_kv_heads: 2, // KV heads (4:1 ratio)
        vocab_size: 100,
        context_length: 1024,
        rope_theta: 10000.0,
        eps: 1e-5,
        rope_type: 0,
            explicit_head_dim: None,
            query_pre_attn_scalar: None,
        bos_token_id: None,
            eos_token_id: None,
    };

    // Create model with dummy weights
    let hidden_dim = config.hidden_dim;
    let head_dim = hidden_dim / config.num_heads; // 4
    let kv_dim = config.num_kv_heads * head_dim; // 2 * 4 = 8

    // Q: [hidden_dim] = [32] - 8 heads
    // K/V: [kv_dim] = [8] - 2 heads
    let q = vec![1.0f32; hidden_dim];
    let current_k = vec![1.0f32; kv_dim];
    let current_v = vec![1.0f32; kv_dim];

    // Empty cache for first position
    let k_cache: Vec<f32> = vec![];
    let v_cache: Vec<f32> = vec![];

    // Test that GQA attention computes correctly
    // Q heads 0-3 should use KV head 0
    // Q heads 4-7 should use KV head 1
    let model = create_test_model_with_config(&config);
    let output = model.attention_with_cache_gqa(&q, &k_cache, &v_cache, &current_k, &current_v);

    // Output should have hidden_dim elements
    assert_eq!(
        output.len(),
        hidden_dim,
        "IMP-105: GQA output should have hidden_dim={hidden_dim} elements"
    );

    // Each head's output should be non-zero (softmax weight = 1.0 for single position)
    for head in 0..config.num_heads {
        let head_start = head * head_dim;
        let head_sum: f32 = output[head_start..head_start + head_dim].iter().sum();
        assert!(
            head_sum.abs() > 1e-6,
            "IMP-105: GQA head {head} output should be non-zero"
        );
    }
}

#[test]
fn test_imp_105_gqa_kv_head_sharing() {
    // IMP-105: Verify that multiple Q heads correctly share KV heads
    let config = GGUFConfig {
        architecture: "test".to_string(),
        constraints: crate::gguf::ArchConstraints::from_architecture("test"),
        hidden_dim: 16, // 4 heads * 4 head_dim
        intermediate_dim: 64,
        num_layers: 1,
        num_heads: 4,    // Q heads
        num_kv_heads: 2, // KV heads (2:1 ratio)
        vocab_size: 100,
        context_length: 1024,
        rope_theta: 10000.0,
        eps: 1e-5,
        rope_type: 0,
            explicit_head_dim: None,
            query_pre_attn_scalar: None,
        bos_token_id: None,
            eos_token_id: None,
    };

    let hidden_dim = config.hidden_dim;
    let head_dim = hidden_dim / config.num_heads; // 4
    let kv_dim = config.num_kv_heads * head_dim; // 8

    // Create Q with different values per head
    let mut q = vec![0.0f32; hidden_dim];
    for head in 0..config.num_heads {
        for d in 0..head_dim {
            q[head * head_dim + d] = (head + 1) as f32;
        }
    }

    // Create K with different values per KV head
    let mut current_k = vec![0.0f32; kv_dim];
    for kv_head in 0..config.num_kv_heads {
        for d in 0..head_dim {
            current_k[kv_head * head_dim + d] = (kv_head + 1) as f32 * 0.5;
        }
    }

    // V values
    let mut current_v = vec![0.0f32; kv_dim];
    for kv_head in 0..config.num_kv_heads {
        for d in 0..head_dim {
            current_v[kv_head * head_dim + d] = (kv_head + 1) as f32;
        }
    }

    let k_cache: Vec<f32> = vec![];
    let v_cache: Vec<f32> = vec![];

    let model = create_test_model_with_config(&config);
    let output = model.attention_with_cache_gqa(&q, &k_cache, &v_cache, &current_k, &current_v);

    // Q heads 0,1 should use KV head 0 (value=1.0)
    // Q heads 2,3 should use KV head 1 (value=2.0)
    // With softmax weight = 1.0 (single position), output = V
    let eps = 1e-5;

    // Head 0 and 1 should have similar outputs (both use KV head 0)
    let head0_sum: f32 = output[0..head_dim].iter().sum();
    let head1_sum: f32 = output[head_dim..2 * head_dim].iter().sum();

    // Head 2 and 3 should have similar outputs (both use KV head 1)
    let head2_sum: f32 = output[2 * head_dim..3 * head_dim].iter().sum();
    let head3_sum: f32 = output[3 * head_dim..4 * head_dim].iter().sum();

    // Verify KV head sharing pattern
    assert!(
        (head0_sum - head1_sum).abs() < eps,
        "IMP-105: Heads 0,1 should produce same output (share KV head 0)"
    );
    assert!(
        (head2_sum - head3_sum).abs() < eps,
        "IMP-105: Heads 2,3 should produce same output (share KV head 1)"
    );
    assert!(
        (head0_sum - head2_sum).abs() > eps,
        "IMP-105: Heads using different KV heads should have different outputs"
    );
}