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
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291


// ============================================================================
// GGUFTransformer: tied embeddings (output.weight missing → uses token_embd)
// ============================================================================

#[test]
fn test_transformer_from_gguf_tied_embeddings() {
    use crate::gguf::GGUFTransformer;

    let vocab = 8;
    let hidden = 4;

    // Note: NO output.weight tensor - should fallback to token_embd.weight
    let data = GGUFBuilder::new()
        .architecture("llama")
        .hidden_dim("llama", hidden as u32)
        .num_layers("llama", 1)
        .num_heads("llama", 1)
        .num_kv_heads("llama", 1)
        .context_length("llama", 32)
        .rope_freq_base("llama", 10000.0)
        .rms_epsilon("llama", 1e-5)
        .ffn_hidden_dim("llama", 8)
        .vocab_size("llama", vocab as u32)
        .add_f32_tensor(
            "token_embd.weight",
            &[hidden as u64, vocab as u64],
            &create_f32_embedding_data(vocab, hidden),
        )
        .add_f32_tensor(
            "blk.0.attn_norm.weight",
            &[hidden as u64],
            &create_f32_norm_weights(hidden),
        )
        .add_f32_tensor(
            "blk.0.attn_q.weight",
            &[hidden as u64, hidden as u64],
            &vec![0.01f32; hidden * hidden],
        )
        .add_f32_tensor(
            "blk.0.attn_k.weight",
            &[hidden as u64, hidden as u64],
            &vec![0.01f32; hidden * hidden],
        )
        .add_f32_tensor(
            "blk.0.attn_v.weight",
            &[hidden as u64, hidden as u64],
            &vec![0.01f32; hidden * hidden],
        )
        .add_f32_tensor(
            "blk.0.attn_output.weight",
            &[hidden as u64, hidden as u64],
            &vec![0.01f32; hidden * hidden],
        )
        .add_f32_tensor(
            "blk.0.ffn_gate.weight",
            &[hidden as u64, 8],
            &vec![0.01f32; hidden * 8],
        )
        .add_f32_tensor(
            "blk.0.ffn_up.weight",
            &[hidden as u64, 8],
            &vec![0.01f32; hidden * 8],
        )
        .add_f32_tensor(
            "blk.0.ffn_down.weight",
            &[8, hidden as u64],
            &vec![0.01f32; 8 * hidden],
        )
        .add_f32_tensor(
            "blk.0.ffn_norm.weight",
            &[hidden as u64],
            &create_f32_norm_weights(hidden),
        )
        .add_f32_tensor(
            "output_norm.weight",
            &[hidden as u64],
            &create_f32_norm_weights(hidden),
        )
        // No output.weight — tied embeddings
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    let transformer = GGUFTransformer::from_gguf(&model, &data);
    assert!(
        transformer.is_ok(),
        "Tied embeddings failed: {:?}",
        transformer.err()
    );
    let t = transformer.expect("test value should be present");
    // lm_head_weight should equal token_embedding (tied)
    assert_eq!(t.lm_head_weight.len(), t.token_embedding.len());
}

// ============================================================================
// GGUFTransformer: 2-layer model
// ============================================================================

#[test]
fn test_transformer_from_gguf_two_layers() {
    use crate::gguf::GGUFTransformer;

    let vocab = 8;
    let hidden = 4;

    let mut builder = GGUFBuilder::new()
        .architecture("llama")
        .hidden_dim("llama", hidden as u32)
        .num_layers("llama", 2)
        .num_heads("llama", 1)
        .num_kv_heads("llama", 1)
        .context_length("llama", 32)
        .rope_freq_base("llama", 10000.0)
        .rms_epsilon("llama", 1e-5)
        .ffn_hidden_dim("llama", 8)
        .vocab_size("llama", vocab as u32)
        .add_f32_tensor(
            "token_embd.weight",
            &[hidden as u64, vocab as u64],
            &create_f32_embedding_data(vocab, hidden),
        );

    // Add both layers
    for layer_idx in 0..2 {
        let prefix = format!("blk.{}", layer_idx);
        builder = builder
            .add_f32_tensor(
                &format!("{}.attn_norm.weight", prefix),
                &[hidden as u64],
                &create_f32_norm_weights(hidden),
            )
            .add_f32_tensor(
                &format!("{}.attn_q.weight", prefix),
                &[hidden as u64, hidden as u64],
                &vec![0.01f32; hidden * hidden],
            )
            .add_f32_tensor(
                &format!("{}.attn_k.weight", prefix),
                &[hidden as u64, hidden as u64],
                &vec![0.01f32; hidden * hidden],
            )
            .add_f32_tensor(
                &format!("{}.attn_v.weight", prefix),
                &[hidden as u64, hidden as u64],
                &vec![0.01f32; hidden * hidden],
            )
            .add_f32_tensor(
                &format!("{}.attn_output.weight", prefix),
                &[hidden as u64, hidden as u64],
                &vec![0.01f32; hidden * hidden],
            )
            .add_f32_tensor(
                &format!("{}.ffn_gate.weight", prefix),
                &[hidden as u64, 8],
                &vec![0.01f32; hidden * 8],
            )
            .add_f32_tensor(
                &format!("{}.ffn_up.weight", prefix),
                &[hidden as u64, 8],
                &vec![0.01f32; hidden * 8],
            )
            .add_f32_tensor(
                &format!("{}.ffn_down.weight", prefix),
                &[8, hidden as u64],
                &vec![0.01f32; 8 * hidden],
            )
            .add_f32_tensor(
                &format!("{}.ffn_norm.weight", prefix),
                &[hidden as u64],
                &create_f32_norm_weights(hidden),
            );
    }

    let data = builder
        .add_f32_tensor(
            "output_norm.weight",
            &[hidden as u64],
            &create_f32_norm_weights(hidden),
        )
        .add_f32_tensor(
            "output.weight",
            &[hidden as u64, vocab as u64],
            &vec![0.01f32; hidden * vocab],
        )
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    let transformer = GGUFTransformer::from_gguf(&model, &data);
    assert!(
        transformer.is_ok(),
        "2-layer failed: {:?}",
        transformer.err()
    );
    let t = transformer.expect("test value should be present");
    assert_eq!(t.layers.len(), 2);
    assert_eq!(t.config.num_layers, 2);
}

// ============================================================================
// Metadata accessor: bos_token_id, eos_token_id
// ============================================================================

#[test]
fn test_bos_eos_token_ids() {
    let data = GGUFBuilder::new()
        .architecture("llama")
        .hidden_dim("llama", 32)
        .num_layers("llama", 1)
        .num_heads("llama", 4)
        .add_u32("tokenizer.ggml.bos_token_id", 1)
        .add_u32("tokenizer.ggml.eos_token_id", 2)
        .build();
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    assert_eq!(model.bos_token_id(), Some(1));
    assert_eq!(model.eos_token_id(), Some(2));
}

#[test]
fn test_bos_eos_missing() {
    let data = GGUFBuilder::new()
        .architecture("llama")
        .hidden_dim("llama", 32)
        .num_layers("llama", 1)
        .num_heads("llama", 4)
        .build();
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    // When not set, should return None (not panic)
    let _ = model.bos_token_id();
    let _ = model.eos_token_id();
}

// ============================================================================
// Metadata: ffn_hidden_dim (intermediate_dim) accessor
// ============================================================================

#[test]
fn test_metadata_ffn_hidden_dim() {
    let data = GGUFBuilder::new()
        .architecture("llama")
        .hidden_dim("llama", 64)
        .num_layers("llama", 1)
        .num_heads("llama", 4)
        .ffn_hidden_dim("llama", 256)
        .build();
    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    // FFN hidden dim is accessed through config when building transformer
    // Just verify the metadata parses correctly
    assert_eq!(model.header.version, 3);
}

// ============================================================================
// Multiple quantization types in one model
// ============================================================================

#[test]
fn test_model_with_mixed_quantization() {
    let data = GGUFBuilder::new()
        .architecture("llama")
        .hidden_dim("llama", 32)
        .num_layers("llama", 1)
        .num_heads("llama", 1)
        .add_f32_tensor("norm.weight", &[32], &vec![1.0f32; 32])
        .add_q4_0_tensor("layer.q4_0", &[32], &create_q4_0_data(32))
        .add_q8_0_tensor("layer.q8_0", &[32], &create_q8_0_data(32))
        .add_q4_k_tensor("layer.q4_k", &[256], &create_q4_k_data(256))
        .add_q5_k_tensor("layer.q5_k", &[256], &create_q5_k_data(256))
        .add_q6_k_tensor("layer.q6_k", &[256], &create_q6_k_data(256))
        .build();

    let model = GGUFModel::from_bytes(&data).expect("test value should be present");
    assert_eq!(model.tensors.len(), 6);

    // Verify each tensor can be dequantized
    let norm = model.get_tensor_f32("norm.weight", &data).expect("test value should be present");
    assert_eq!(norm.len(), 32);

    let q4_0 = model.get_tensor_f32("layer.q4_0", &data).expect("test value should be present");
    assert_eq!(q4_0.len(), 32);

    let q8_0 = model.get_tensor_f32("layer.q8_0", &data).expect("test value should be present");
    assert_eq!(q8_0.len(), 32);

    let q4_k = model.get_tensor_f32("layer.q4_k", &data).expect("test value should be present");
    assert_eq!(q4_k.len(), 256);

    let q5_k = model.get_tensor_f32("layer.q5_k", &data).expect("test value should be present");
    assert_eq!(q5_k.len(), 256);

    let q6_k = model.get_tensor_f32("layer.q6_k", &data).expect("test value should be present");
    assert_eq!(q6_k.len(), 256);
}