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

//! Convert Module Tests Part 03 - T-COV-95 Deep Coverage Bridge
//!
//! Tests for:
//! - GgufToAprQ4KConverter::infer_rope_type for all architecture families
//! - GgufToAprQ4KConverter helper methods: get_string, get_u32, get_f32
//! - RawTensor struct construction and fields
//! - Q4KConversionStats Debug/Clone
//! - ConversionStats edge cases (zero values)
//! - to_apr_bytes / from_apr_bytes additional error paths
//!
//! Refs PMAT-802: Protocol T-COV-95

#[cfg(test)]
mod tests {
    use crate::apr_transformer::{AprTransformer, AprTransformerConfig, AprTransformerLayer};
    use crate::convert::*;
    use std::collections::HashMap;

    #[test]
    fn test_to_apr_bytes_roundtrip_with_biases() {
        let config = AprTransformerConfig {
            architecture: "phi2".to_string(),
            hidden_dim: 4,
            num_layers: 1,
            num_heads: 2,
            num_kv_heads: 2,
            vocab_size: 8,
            intermediate_dim: 8,
            context_length: 64,
            rope_theta: 10000.0,
            eps: 1e-5,
            eos_token_id: None,
        ..Default::default()
        };

        let hidden = config.hidden_dim;
        let vocab = config.vocab_size;
        let intermediate = config.intermediate_dim;

        let transformer = AprTransformer {
            config,
            token_embedding: vec![0.1f32; vocab * hidden],
            layers: vec![AprTransformerLayer {
                attn_norm_weight: vec![1.0; hidden],
                attn_norm_bias: Some(vec![0.0; hidden]), // phi2 has biases
                qkv_weight: vec![0.01; hidden * hidden * 3],
                qkv_bias: Some(vec![0.0; hidden * 3]),
                attn_output_weight: vec![0.01; hidden * hidden],
                attn_output_bias: Some(vec![0.0; hidden]),
                ffn_gate_weight: None, // phi2 has no gate
                ffn_gate_bias: None,
                ffn_up_weight: vec![0.01; hidden * intermediate],
                ffn_up_bias: Some(vec![0.0; intermediate]),
                ffn_down_weight: vec![0.01; intermediate * hidden],
                ffn_down_bias: Some(vec![0.0; hidden]),
                ffn_norm_weight: Some(vec![1.0; hidden]),
                ffn_norm_bias: Some(vec![0.0; hidden]),
                attn_q_norm_weight: None,
                attn_k_norm_weight: None,
                linear_attn_z_weight: None,
                linear_attn_b_weight: None,
                linear_attn_a_weight: None,
                linear_attn_conv1d_weight: None,
                linear_attn_a_log: None,
                linear_attn_dt_bias: None,
                linear_attn_norm_weight: None,
                moe_gate_weight: None,
                moe_expert_gate_up: None,
                moe_expert_down: None,
                moe_shared_gate: None,
                moe_shared_up: None,
                moe_shared_down: None,
                moe_shared_expert_gate_weight: None,
            }],
            output_norm_weight: vec![1.0; hidden],
            output_norm_bias: Some(vec![0.0; hidden]),
            lm_head_weight: vec![0.01; hidden * vocab],
            lm_head_bias: Some(vec![0.0; vocab]),
            q4k_layers: None,
            lm_head_weight_q6k: None,
            lm_head_weight_q4k: None,
        };

        let bytes = GgufToAprConverter::to_apr_bytes(&transformer).unwrap();
        let restored = GgufToAprConverter::from_apr_bytes(&bytes).unwrap();
        assert_eq!(restored.config.architecture, "phi2");
        assert_eq!(restored.config.num_heads, 2);
        // Verify biases survived roundtrip
        assert!(restored.layers[0].attn_norm_bias.is_some());
        assert!(restored.layers[0].qkv_bias.is_some());
    }

    // =========================================================================
    // BUG-APR-002: Byte size calculation for quantized tensors
    // =========================================================================
    // The Q4K converter was using integer division (/) which rounds DOWN,
    // causing "tensor exceeds file bounds" for tensors not perfectly divisible
    // by block size. Fix: use div_ceil() to round UP.

    #[test]
    fn test_bug_apr_002_q4k_byte_size_div_ceil() {
        // Q4_K block: 256 elements = 144 bytes
        // For 65600 elements (256 * 256 + 64):
        // - Wrong: (65600 / 256) * 144 = 256 * 144 = 36864 bytes
        // - Right: 65600.div_ceil(256) * 144 = 257 * 144 = 37008 bytes
        let num_elements = 65600usize;
        let wrong_byte_size = (num_elements / 256) * 144;
        let right_byte_size = num_elements.div_ceil(256) * 144;

        assert_eq!(
            wrong_byte_size, 36864,
            "Wrong calculation (integer division)"
        );
        assert_eq!(right_byte_size, 37008, "Correct calculation (div_ceil)");
        assert!(
            right_byte_size > wrong_byte_size,
            "div_ceil must give larger result"
        );
    }

    #[test]
    fn test_bug_apr_002_q8_byte_size_div_ceil() {
        // Q8_0 block: 32 elements = 34 bytes (2 scale + 32 quants)
        // For 1000 elements (32 * 31 + 8):
        // - Wrong: (1000 / 32) * 34 = 31 * 34 = 1054 bytes
        // - Right: 1000.div_ceil(32) * 34 = 32 * 34 = 1088 bytes
        let num_elements = 1000usize;
        let wrong_byte_size = (num_elements / 32) * 34;
        let right_byte_size = num_elements.div_ceil(32) * 34;

        assert_eq!(
            wrong_byte_size, 1054,
            "Wrong calculation (integer division)"
        );
        assert_eq!(right_byte_size, 1088, "Correct calculation (div_ceil)");
        assert!(
            right_byte_size > wrong_byte_size,
            "div_ceil must give larger result"
        );
    }

    #[test]
    fn test_bug_apr_002_exact_divisibility_no_change() {
        // When num_elements is perfectly divisible, both methods give same result
        let num_elements = 65536usize; // 256 * 256
        let old_style = (num_elements / 256) * 144;
        let new_style = num_elements.div_ceil(256) * 144;

        assert_eq!(
            old_style, new_style,
            "Exact divisibility should give same result"
        );
        assert_eq!(new_style, 36864);
    }

    #[test]
    fn test_bug_apr_002_q5k_byte_size_div_ceil() {
        // Q5_K block: 256 elements = 176 bytes
        // For 65537 elements (256 * 256 + 1):
        // - Wrong: (65537 / 256) * 176 = 256 * 176 = 45056 bytes
        // - Right: 65537.div_ceil(256) * 176 = 257 * 176 = 45232 bytes
        let num_elements = 65537usize;
        let wrong_byte_size = (num_elements / 256) * 176;
        let right_byte_size = num_elements.div_ceil(256) * 176;

        assert_ne!(wrong_byte_size, right_byte_size);
        assert_eq!(right_byte_size - wrong_byte_size, 176); // One extra block
    }

    #[test]
    fn test_bug_apr_002_q6k_byte_size_div_ceil() {
        // Q6_K block: 256 elements = 210 bytes
        // For 65600 elements:
        // - Wrong: (65600 / 256) * 210 = 256 * 210 = 53760 bytes
        // - Right: 65600.div_ceil(256) * 210 = 257 * 210 = 53970 bytes
        let num_elements = 65600usize;
        let wrong_byte_size = (num_elements / 256) * 210;
        let right_byte_size = num_elements.div_ceil(256) * 210;

        assert_ne!(wrong_byte_size, right_byte_size);
        assert_eq!(right_byte_size - wrong_byte_size, 210); // One extra block
    }
include!("tests_infer_rope.rs");
}