realizar 0.8.4

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

//! Tests for flash_attention_tiled (PARITY-026)
//!
//! Coverage target: src/gguf/inference/attention.rs flash_attention_tiled (73 uncov, 0%)

use crate::gguf::test_helpers::create_test_model_with_config;
use crate::gguf::GGUFConfig;

fn small_config() -> GGUFConfig {
    GGUFConfig {
        architecture: "llama".to_string(),
        constraints: crate::gguf::ArchConstraints::from_architecture("llama"),
        hidden_dim: 64,
        intermediate_dim: 128,
        num_heads: 4,
        num_kv_heads: 4,
        num_layers: 1,
        vocab_size: 100,
        context_length: 256,
        rope_theta: 10000.0,
        eps: 1e-5,
        rope_type: 0,
        explicit_head_dim: None,
        bos_token_id: None,
        eos_token_id: None,
    }
}

#[test]
fn test_flash_attention_tiled_no_cache() {
    let model = create_test_model_with_config(&small_config());
    let hidden_dim = model.config.hidden_dim;

    // Query: single position
    let q = vec![1.0f32; hidden_dim];
    let k_cache: Vec<f32> = vec![]; // No cache
    let v_cache: Vec<f32> = vec![];
    let current_k = vec![0.5f32; hidden_dim];
    let current_v = vec![0.3f32; hidden_dim];

    let output = model.flash_attention_tiled(&q, &k_cache, &v_cache, &current_k, &current_v, 64);

    assert_eq!(output.len(), hidden_dim);
    // With no cache, each head attends only to current — output should be current_v per head
    assert!(output.iter().all(|x| x.is_finite()));
    // Since there's only one KV pair (current), attention output should match current_v
    for (o, &expected) in output.iter().zip(current_v.iter()) {
        let diff = (o - expected).abs();
        assert!(diff < 1e-3, "output={o} expected={expected} diff={diff}");
    }
}

#[test]
fn test_flash_attention_tiled_with_cache() {
    let model = create_test_model_with_config(&small_config());
    let hidden_dim = model.config.hidden_dim;

    // Query
    let q = vec![1.0f32; hidden_dim];

    // Cache with 3 previous positions
    let cache_len = 3;
    let k_cache = vec![0.1f32; cache_len * hidden_dim];
    let v_cache = vec![0.2f32; cache_len * hidden_dim];

    let current_k = vec![1.0f32; hidden_dim];
    let current_v = vec![0.5f32; hidden_dim];

    let output = model.flash_attention_tiled(&q, &k_cache, &v_cache, &current_k, &current_v, 64);

    assert_eq!(output.len(), hidden_dim);
    assert!(output.iter().all(|x| x.is_finite()));
    // Output should be a softmax-weighted average of v_cache entries + current_v
    // With uniform-ish queries, should be between 0.2 and 0.5
    for val in &output {
        assert!(*val > 0.0 && *val < 1.0, "val={val} out of expected range");
    }
}

#[test]
fn test_flash_attention_tiled_multi_tile() {
    let model = create_test_model_with_config(&small_config());
    let hidden_dim = model.config.hidden_dim;

    // Use small block_size=2 to force multiple tiles
    let cache_len = 5;
    let k_cache = vec![0.1f32; cache_len * hidden_dim];
    let v_cache = vec![0.2f32; cache_len * hidden_dim];
    let q = vec![1.0f32; hidden_dim];
    let current_k = vec![0.5f32; hidden_dim];
    let current_v = vec![0.3f32; hidden_dim];

    // block_size=2, total_len=6, so 3 tiles
    let output = model.flash_attention_tiled(&q, &k_cache, &v_cache, &current_k, &current_v, 2);

    assert_eq!(output.len(), hidden_dim);
    assert!(output.iter().all(|x| x.is_finite()));
}

#[test]
fn test_flash_attention_tiled_parity_with_standard() {
    let model = create_test_model_with_config(&small_config());
    let hidden_dim = model.config.hidden_dim;

    let q = vec![1.0f32; hidden_dim];
    let cache_len = 2;
    let k_cache = vec![0.3f32; cache_len * hidden_dim];
    let v_cache = vec![0.4f32; cache_len * hidden_dim];
    let current_k = vec![0.7f32; hidden_dim];
    let current_v = vec![0.6f32; hidden_dim];

    // Flash attention result
    let flash_output =
        model.flash_attention_tiled(&q, &k_cache, &v_cache, &current_k, &current_v, 64);

    // Standard attention result
    let standard_output =
        model.attention_with_cache(&q, &k_cache, &v_cache, &current_k, &current_v);

    assert_eq!(flash_output.len(), standard_output.len());
    for (f, s) in flash_output.iter().zip(standard_output.iter()) {
        let diff = (f - s).abs();
        assert!(
            diff < 1e-3,
            "flash={f} standard={s} diff={diff} — parity failure"
        );
    }
}

#[test]
fn test_flash_attention_tiled_large_cache() {
    let model = create_test_model_with_config(&small_config());
    let hidden_dim = model.config.hidden_dim;

    // Large cache with 100 positions, block_size=32
    let cache_len = 100;
    let k_cache: Vec<f32> = (0..cache_len * hidden_dim)
        .map(|i| ((i % 7) as f32) * 0.1)
        .collect();
    let v_cache: Vec<f32> = (0..cache_len * hidden_dim)
        .map(|i| ((i % 11) as f32) * 0.05)
        .collect();

    let q = vec![0.5f32; hidden_dim];
    let current_k = vec![0.3f32; hidden_dim];
    let current_v = vec![0.2f32; hidden_dim];

    let output = model.flash_attention_tiled(&q, &k_cache, &v_cache, &current_k, &current_v, 32);

    assert_eq!(output.len(), hidden_dim);
    assert!(output.iter().all(|x| x.is_finite()));
}

#[test]
fn test_flash_attention_tiled_gqa() {
    let gqa_config = GGUFConfig {
        architecture: "llama".to_string(),
        constraints: crate::gguf::ArchConstraints::from_architecture("llama"),
        hidden_dim: 64,
        intermediate_dim: 128,
        num_heads: 8,
        num_kv_heads: 2,
        num_layers: 1,
        vocab_size: 100,
        context_length: 256,
        rope_theta: 10000.0,
        eps: 1e-5,
        rope_type: 0,
        explicit_head_dim: None,
        bos_token_id: None,
        eos_token_id: None,
    };

    let model = create_test_model_with_config(&gqa_config);
    let hidden_dim = model.config.hidden_dim;

    let q = vec![1.0f32; hidden_dim];
    let k_cache = vec![0.2f32; 3 * hidden_dim];
    let v_cache = vec![0.3f32; 3 * hidden_dim];
    let current_k = vec![0.5f32; hidden_dim];
    let current_v = vec![0.4f32; hidden_dim];

    let output = model.flash_attention_tiled(&q, &k_cache, &v_cache, &current_k, &current_v, 64);

    assert_eq!(output.len(), hidden_dim);
    assert!(output.iter().all(|x| x.is_finite()));
}