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
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


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

    #[test]
    fn test_owned_quantized_model_debug_shows_all_fields() {
        use super::super::quantized::OwnedQuantizedTensor;
        use super::super::types::GGUF_TYPE_Q8_0;

        let config = GGUFConfig {
            architecture: "llama_debug".to_string(),
            constraints: crate::gguf::ArchConstraints::from_architecture("llama_debug"),
            hidden_dim: 128,
            num_layers: 4,
            num_heads: 4,
            num_kv_heads: 4,
            vocab_size: 1000,
            intermediate_dim: 256,
            context_length: 512,
            rope_theta: 10000.0,
            eps: 1e-5,
            rope_type: 0,
            explicit_head_dim: None,
            bos_token_id: None,
            eos_token_id: None,
        };

        let model = OwnedQuantizedModel {
            config,
            token_embedding: vec![0.0; 128000],
            position_embedding: None,
            layers: vec![],
            encoder_layers: vec![],
            encoder_output_norm_weight: None,
            encoder_output_norm_bias: None,
            output_norm_weight: vec![1.0; 128],
            output_norm_bias: None,
            lm_head_weight: OwnedQuantizedTensor {
                data: vec![0u8; 1000],
                in_dim: 128,
                out_dim: 1000,
                qtype: GGUF_TYPE_Q8_0,
            },
            lm_head_bias: None,
            #[cfg(feature = "cuda")]
            cuda_executor: None,
            #[cfg(feature = "cuda")]
            cuda_kernel_count: std::sync::atomic::AtomicU64::new(42),
            #[cfg(feature = "cuda")]
            cached_weight_names: std::sync::Mutex::new(std::collections::HashSet::new()),
        };

        let debug_str = format!("{:?}", model);
        assert!(debug_str.contains("llama_debug"));
        assert!(debug_str.contains("token_embedding_len"));
        assert!(debug_str.contains("128000"));
        assert!(debug_str.contains("layers_count"));
        assert!(debug_str.contains("output_norm_weight_len"));
        assert!(debug_str.contains("has_output_norm_bias"));
        assert!(debug_str.contains("lm_head_weight"));
        assert!(debug_str.contains("has_lm_head_bias"));
        #[cfg(feature = "cuda")]
        assert!(debug_str.contains("cuda_enabled"));
    }

    #[test]
    fn test_gguf_config_fields() {
        let config = GGUFConfig {
            architecture: "mistral".to_string(),
            constraints: crate::gguf::ArchConstraints::from_architecture("mistral"),
            hidden_dim: 4096,
            num_layers: 32,
            num_heads: 32,
            num_kv_heads: 8, // GQA style
            vocab_size: 32000,
            intermediate_dim: 14336,
            context_length: 8192,
            rope_theta: 1000000.0,
            eps: 1e-5,
            rope_type: 1,
            explicit_head_dim: None,
            bos_token_id: None,
            eos_token_id: None,
        };

        assert_eq!(config.architecture, "mistral");
        assert_eq!(config.hidden_dim, 4096);
        assert_eq!(config.num_layers, 32);
        assert_eq!(config.num_heads, 32);
        assert_eq!(config.num_kv_heads, 8);
        assert_eq!(config.vocab_size, 32000);
        assert_eq!(config.intermediate_dim, 14336);
        assert_eq!(config.context_length, 8192);
        assert!((config.rope_theta - 1000000.0).abs() < 0.1);
        assert!((config.eps - 1e-5).abs() < 1e-10);
        assert_eq!(config.rope_type, 1);
    }

    #[test]
    fn test_mapped_gguf_model_file_not_found() {
        let result = MappedGGUFModel::from_path("/nonexistent/path/model.gguf");
        assert!(result.is_err());
    }

    // =========================================================================
    // GH-40 FALSIFICATION: OwnedQuantizedModel API contract stability
    // =========================================================================

    /// GH-40: OwnedQuantizedModel must expose all public fields required by
    /// the inference pipeline. If a field is renamed, removed, or made private,
    /// this test fails — catching API breakage at compile time.
    #[test]
    fn test_falsify_gh40_api_contract_all_fields_accessible() {
        use super::super::quantized::OwnedQuantizedTensor;
        use super::super::types::GGUF_TYPE_Q4_K;

        let config = GGUFConfig {
            architecture: "gh40_test".to_string(),
            constraints: crate::gguf::ArchConstraints::from_architecture("gh40_test"),
            hidden_dim: 64,
            num_layers: 1,
            num_heads: 2,
            num_kv_heads: 2,
            vocab_size: 100,
            intermediate_dim: 128,
            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 = OwnedQuantizedModel {
            config,
            token_embedding: vec![0.1; 6400],
            position_embedding: None,
            layers: vec![],
            encoder_layers: vec![],
            encoder_output_norm_weight: None,
            encoder_output_norm_bias: None,
            output_norm_weight: vec![1.0; 64],
            output_norm_bias: None,
            lm_head_weight: OwnedQuantizedTensor {
                data: vec![0u8; 128],
                in_dim: 64,
                out_dim: 100,
                qtype: GGUF_TYPE_Q4_K,
            },
            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()),
        };

        // GH-40 contract: every field below must be pub and have the expected type.
        // If ANY of these lines fails to compile, the API contract is broken.
        let _arch: &str = &model.config.architecture;
        let _hidden: usize = model.config.hidden_dim;
        let _embed: &[f32] = &model.token_embedding;
        let _layers: &[super::super::quantized::OwnedQuantizedLayer] = &model.layers;
        let _norm: &[f32] = &model.output_norm_weight;
        let _norm_bias: &Option<Vec<f32>> = &model.output_norm_bias;
        let _lm_data: &[u8] = &model.lm_head_weight.data;
        let _lm_in: usize = model.lm_head_weight.in_dim;
        let _lm_out: usize = model.lm_head_weight.out_dim;
        let _lm_qt: u32 = model.lm_head_weight.qtype;
        let _lm_bias: &Option<Vec<f32>> = &model.lm_head_bias;

        // GH-40 contract: config fields required by inference
        assert_eq!(model.config.hidden_dim, 64);
        assert_eq!(model.config.num_heads, 2);
        assert_eq!(model.config.num_kv_heads, 2);
        assert_eq!(model.config.vocab_size, 100);
        assert_eq!(model.config.intermediate_dim, 128);
        assert!(model.config.rope_theta > 0.0);
        assert!(model.config.eps > 0.0);
    }

    /// GH-40: GGUFConfig must have bos_token_id field (added in the fix).
    /// Without this field, tokenizer initialization fails silently.
    #[test]
    fn test_falsify_gh40_config_has_bos_token_id() {
        let config = GGUFConfig {
            architecture: "test".to_string(),
            constraints: crate::gguf::ArchConstraints::from_architecture("test"),
            hidden_dim: 64,
            num_layers: 1,
            num_heads: 1,
            num_kv_heads: 1,
            vocab_size: 100,
            intermediate_dim: 128,
            context_length: 256,
            rope_theta: 10000.0,
            eps: 1e-5,
            rope_type: 0,
            explicit_head_dim: None,
            bos_token_id: Some(1),
            eos_token_id: Some(2),
        };
        assert_eq!(
            config.bos_token_id,
            Some(1),
            "GH-40: bos_token_id must be accessible"
        );
    }
include!("model_gguf_transformer.rs");
}