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
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


#[cfg(test)]
mod tests {
    use super::*;
    use crate::gguf::test_factory::{build_minimal_llama_gguf, build_minimal_phi2_gguf};

    // ============================================================================
    // QuantizedGGUFTransformerLayer tests
    // ============================================================================

    #[test]
    fn test_quantized_layer_struct_fields() {
        // Verify struct field layout and optional fields
        let layer = QuantizedGGUFTransformerLayer {
            attn_norm_weight: vec![1.0; 64],
            attn_norm_bias: None,
            qkv_weight: QKVWeights::Fused(QuantizedTensorRef {
                offset: 0,
                byte_size: 100,
                num_elements: 192,
                qtype: GGUF_TYPE_Q4_K,
            }),
            qkv_bias: None,
            attn_output_weight: QuantizedTensorRef {
                offset: 0,
                byte_size: 50,
                num_elements: 64,
                qtype: GGUF_TYPE_Q4_K,
            },
            attn_output_bias: None,
            ffn_up_weight: QuantizedTensorRef {
                offset: 0,
                byte_size: 100,
                num_elements: 256,
                qtype: GGUF_TYPE_Q4_K,
            },
            ffn_up_bias: None,
            ffn_down_weight: QuantizedTensorRef {
                offset: 0,
                byte_size: 100,
                num_elements: 64,
                qtype: GGUF_TYPE_Q4_K,
            },
            ffn_down_bias: None,
            ffn_gate_weight: None,
            ffn_gate_bias: None,
            ffn_norm_weight: None,
            ffn_norm_bias: None,
            attn_q_norm_weight: None,
            attn_k_norm_weight: None,
        };

        assert_eq!(layer.attn_norm_weight.len(), 64);
        assert!(layer.attn_norm_bias.is_none());
        assert!(layer.ffn_gate_weight.is_none());
    }

    #[test]
    fn test_quantized_layer_with_optional_fields() {
        let layer = QuantizedGGUFTransformerLayer {
            attn_norm_weight: vec![1.0; 32],
            attn_norm_bias: Some(vec![0.0; 32]),
            qkv_weight: QKVWeights::Separate {
                q: QuantizedTensorRef {
                    offset: 0,
                    byte_size: 50,
                    num_elements: 32,
                    qtype: GGUF_TYPE_Q4_0,
                },
                k: QuantizedTensorRef {
                    offset: 50,
                    byte_size: 50,
                    num_elements: 32,
                    qtype: GGUF_TYPE_Q4_0,
                },
                v: QuantizedTensorRef {
                    offset: 100,
                    byte_size: 50,
                    num_elements: 32,
                    qtype: GGUF_TYPE_Q4_0,
                },
            },
            qkv_bias: Some(vec![0.0; 96]),
            attn_output_weight: QuantizedTensorRef {
                offset: 0,
                byte_size: 50,
                num_elements: 32,
                qtype: GGUF_TYPE_Q4_0,
            },
            attn_output_bias: Some(vec![0.0; 32]),
            ffn_up_weight: QuantizedTensorRef {
                offset: 0,
                byte_size: 100,
                num_elements: 128,
                qtype: GGUF_TYPE_Q4_0,
            },
            ffn_up_bias: Some(vec![0.0; 128]),
            ffn_down_weight: QuantizedTensorRef {
                offset: 0,
                byte_size: 50,
                num_elements: 32,
                qtype: GGUF_TYPE_Q4_0,
            },
            ffn_down_bias: Some(vec![0.0; 32]),
            ffn_gate_weight: Some(QuantizedTensorRef {
                offset: 0,
                byte_size: 100,
                num_elements: 128,
                qtype: GGUF_TYPE_Q4_0,
            }),
            ffn_gate_bias: Some(vec![0.0; 128]),
            ffn_norm_weight: Some(vec![1.0; 32]),
            ffn_norm_bias: Some(vec![0.0; 32]),
            attn_q_norm_weight: None,
            attn_k_norm_weight: None,
        };

        assert!(layer.attn_norm_bias.is_some());
        assert!(layer.qkv_bias.is_some());
        assert!(layer.ffn_gate_weight.is_some());
        assert!(layer.ffn_norm_weight.is_some());
    }

    // ============================================================================
    // QuantizedGGUFTransformer tests using test_factory
    // ============================================================================

    #[test]
    fn test_quantized_transformer_from_llama_gguf() {
        // build_minimal_llama_gguf(vocab_size, hidden_dim, intermediate_dim, num_heads, num_kv_heads)
        let gguf_data = build_minimal_llama_gguf(100, 64, 256, 4, 4);
        let model = GGUFModel::from_bytes(&gguf_data).expect("Failed to parse GGUF");
        let transformer = QuantizedGGUFTransformer::from_gguf(&model, &gguf_data);

        assert!(
            transformer.is_ok(),
            "Failed to load transformer: {:?}",
            transformer.err()
        );
        let transformer = transformer.expect("transformer");

        assert_eq!(transformer.config.hidden_dim, 64);
        assert_eq!(transformer.config.num_layers, 1); // test_factory builds 1 layer
        assert_eq!(transformer.layers.len(), 1);
        assert_eq!(transformer.token_embedding.len(), 100 * 64);
    }

    #[test]
    fn test_quantized_transformer_from_phi2_gguf() {
        // build_minimal_phi2_gguf(vocab_size, hidden_dim, intermediate_dim, num_heads)
        let gguf_data = build_minimal_phi2_gguf(100, 64, 256, 4);
        let model = GGUFModel::from_bytes(&gguf_data).expect("Failed to parse GGUF");
        let transformer = QuantizedGGUFTransformer::from_gguf(&model, &gguf_data);

        assert!(
            transformer.is_ok(),
            "Failed to load transformer: {:?}",
            transformer.err()
        );
        let transformer = transformer.expect("transformer");

        assert_eq!(transformer.config.hidden_dim, 64);
        assert_eq!(transformer.layers.len(), 1);
    }

    #[test]
    fn test_quantized_transformer_config() {
        let gguf_data = build_minimal_llama_gguf(200, 128, 512, 8, 4);
        let model = GGUFModel::from_bytes(&gguf_data).expect("Failed to parse GGUF");
        let transformer = QuantizedGGUFTransformer::from_gguf(&model, &gguf_data).expect("transformer");

        assert_eq!(transformer.config.hidden_dim, 128);
        // intermediate_dim is inferred from ffn_up tensor, may differ from input
        assert!(transformer.config.intermediate_dim > 0);
        assert_eq!(transformer.config.num_heads, 8);
        assert_eq!(transformer.config.num_kv_heads, 4);
        assert_eq!(transformer.config.vocab_size, 200);
    }

    #[test]
    fn test_quantized_transformer_output_norm() {
        let gguf_data = build_minimal_llama_gguf(100, 64, 256, 4, 4);
        let model = GGUFModel::from_bytes(&gguf_data).expect("Failed to parse GGUF");
        let transformer = QuantizedGGUFTransformer::from_gguf(&model, &gguf_data).expect("transformer");

        // Output norm weight should be hidden_dim size
        assert_eq!(transformer.output_norm_weight.len(), 64);
    }

    #[test]
    fn test_quantized_transformer_lm_head() {
        let gguf_data = build_minimal_llama_gguf(100, 64, 256, 4, 4);
        let model = GGUFModel::from_bytes(&gguf_data).expect("Failed to parse GGUF");
        let transformer = QuantizedGGUFTransformer::from_gguf(&model, &gguf_data).expect("transformer");

        // LM head should have elements (may use tied embeddings with F32 type 0)
        assert!(transformer.lm_head_weight.num_elements > 0);
        // qtype can be 0 (F32) for tied embeddings
        assert!(transformer.lm_head_weight.byte_size > 0);
    }

    #[test]
    fn test_quantized_transformer_layer_attn_norm() {
        let gguf_data = build_minimal_llama_gguf(100, 64, 256, 4, 4);
        let model = GGUFModel::from_bytes(&gguf_data).expect("Failed to parse GGUF");
        let transformer = QuantizedGGUFTransformer::from_gguf(&model, &gguf_data).expect("transformer");

        let layer = &transformer.layers[0];
        assert_eq!(layer.attn_norm_weight.len(), 64);
    }

    #[test]
    fn test_quantized_transformer_layer_qkv() {
        let gguf_data = build_minimal_llama_gguf(100, 64, 256, 4, 4);
        let model = GGUFModel::from_bytes(&gguf_data).expect("Failed to parse GGUF");
        let transformer = QuantizedGGUFTransformer::from_gguf(&model, &gguf_data).expect("transformer");

        let layer = &transformer.layers[0];
        // QKV should be either fused or separate
        match &layer.qkv_weight {
            QKVWeights::Fused(ref tensor) => {
                assert!(tensor.num_elements > 0);
            },
            QKVWeights::Separate { q, k, v } => {
                assert!(q.num_elements > 0);
                assert!(k.num_elements > 0);
                assert!(v.num_elements > 0);
            },
        }
    }

    #[test]
    fn test_quantized_transformer_layer_ffn() {
        let gguf_data = build_minimal_llama_gguf(100, 64, 256, 4, 4);
        let model = GGUFModel::from_bytes(&gguf_data).expect("Failed to parse GGUF");
        let transformer = QuantizedGGUFTransformer::from_gguf(&model, &gguf_data).expect("transformer");

        let layer = &transformer.layers[0];
        assert!(layer.ffn_up_weight.num_elements > 0);
        assert!(layer.ffn_down_weight.num_elements > 0);
    }

    #[test]
    fn test_quantized_transformer_has_data_ref() {
        let gguf_data = build_minimal_llama_gguf(100, 64, 256, 4, 4);
        let model = GGUFModel::from_bytes(&gguf_data).expect("Failed to parse GGUF");
        let transformer = QuantizedGGUFTransformer::from_gguf(&model, &gguf_data).expect("transformer");

        // The data reference should point to the original data
        assert!(!transformer.data.is_empty());
        assert_eq!(transformer.data.len(), gguf_data.len());
    }
}