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
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
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378


#[test]
fn test_loader_tensor_info_large_dimension() {
    let mut data = build_header(1, 0);
    // Single very large dimension
    data.extend(build_tensor_info("large", &[1_000_000_000], GGUF_TYPE_F32, 0));

    let model = GGUFModel::from_bytes(&data).expect("Should parse large dim");
    assert_eq!(model.tensors[0].dims[0], 1_000_000_000);
}

#[test]
fn test_loader_tensor_info_multiple_qtypes() {
    let mut data = build_header(5, 0);

    data.extend(build_tensor_info("f32", &[32], GGUF_TYPE_F32, 0));
    data.extend(build_tensor_info("f16", &[32], GGUF_TYPE_F16, 128));
    data.extend(build_tensor_info("q4_0", &[32], GGUF_TYPE_Q4_0, 192));
    data.extend(build_tensor_info("q8_0", &[32], GGUF_TYPE_Q8_0, 210));
    data.extend(build_tensor_info("q4_k", &[256], GGUF_TYPE_Q4_K, 244));

    let model = GGUFModel::from_bytes(&data).expect("Should parse multiple qtypes");

    assert_eq!(model.tensors.len(), 5);
    assert_eq!(model.tensors[0].qtype, GGUF_TYPE_F32);
    assert_eq!(model.tensors[1].qtype, GGUF_TYPE_F16);
    assert_eq!(model.tensors[2].qtype, GGUF_TYPE_Q4_0);
    assert_eq!(model.tensors[3].qtype, GGUF_TYPE_Q8_0);
    assert_eq!(model.tensors[4].qtype, GGUF_TYPE_Q4_K);
}

// =============================================================================
// Alignment Calculation Tests
// =============================================================================

#[test]
fn test_loader_alignment_at_boundary() {
    // Header + metadata that lands exactly on 32-byte boundary
    let mut data = build_header(0, 0);
    assert_eq!(data.len(), 24);

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    // tensor_data_start should be aligned to 32
    assert_eq!(model.tensor_data_start % GGUF_ALIGNMENT, 0);
    assert_eq!(model.tensor_data_start, 32); // 24 rounded up to 32
}

#[test]
fn test_loader_alignment_with_metadata() {
    let mut data = build_header(0, 1);
    // Add metadata that changes alignment requirement
    data.extend(build_metadata("test", 4, &42u32.to_le_bytes()));

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    assert_eq!(model.tensor_data_start % GGUF_ALIGNMENT, 0);
}

#[test]
fn test_loader_alignment_with_tensors() {
    let mut data = build_header(2, 0);
    data.extend(build_tensor_info("t1", &[64], GGUF_TYPE_F32, 0));
    data.extend(build_tensor_info("t2", &[64], GGUF_TYPE_F32, 256));

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    assert_eq!(model.tensor_data_start % GGUF_ALIGNMENT, 0);
}

// =============================================================================
// get_tensor_f32 - Data Range Validation
// =============================================================================

#[test]
fn test_loader_get_tensor_f32_exact_boundary() {
    // Tensor data ends exactly at file end
    let mut data = build_header(1, 0);
    data.extend(build_tensor_info("test", &[4], GGUF_TYPE_F32, 0));
    align_to_boundary(&mut data);

    let tensor_start = data.len();
    // 4 f32 values = 16 bytes
    for i in 0..4 {
        data.extend_from_slice(&(i as f32).to_le_bytes());
    }

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    let tensor = model.get_tensor_f32("test", &data).expect("Should extract");
    assert_eq!(tensor.len(), 4);
}

#[test]
fn test_loader_get_tensor_f32_one_byte_short() {
    // Tensor data is 1 byte short
    let mut data = build_header(1, 0);
    data.extend(build_tensor_info("test", &[4], GGUF_TYPE_F32, 0));
    align_to_boundary(&mut data);

    // 4 f32 = 16 bytes, but only provide 15
    data.extend(&[0u8; 15]);

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    let result = model.get_tensor_f32("test", &data);
    assert!(result.is_err());
}

#[test]
fn test_loader_get_tensor_f32_2d_matrix() {
    let mut data = build_header(1, 0);
    data.extend(build_tensor_info("matrix", &[3, 4], GGUF_TYPE_F32, 0));
    align_to_boundary(&mut data);

    // 3 * 4 = 12 f32 values = 48 bytes
    for i in 0..12 {
        data.extend_from_slice(&(i as f32).to_le_bytes());
    }

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    let tensor = model.get_tensor_f32("matrix", &data).expect("Should extract");
    assert_eq!(tensor.len(), 12);
    assert!((tensor[5] - 5.0).abs() < f32::EPSILON);
}

#[test]
fn test_loader_get_tensor_q4_0_partial_block() {
    // Q4_0 with elements not divisible by block size (32)
    // Tests the truncation after dequantization
    let mut data = build_header(1, 0);
    data.extend(build_tensor_info("test", &[48], GGUF_TYPE_Q4_0, 0)); // 48 = 1.5 blocks
    align_to_boundary(&mut data);

    // Need 2 blocks (64 elements worth) for 48 elements
    // Each Q4_0 block: 2 bytes f16 scale + 16 bytes quants = 18 bytes
    for _ in 0..2 {
        let scale = half::f16::from_f32(1.0);
        data.extend_from_slice(&scale.to_le_bytes());
        data.extend([0x00u8; 16]);
    }

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    let tensor = model.get_tensor_f32("test", &data).expect("Should extract");
    assert_eq!(tensor.len(), 48); // Truncated to exact size
}

#[test]
fn test_loader_get_tensor_q8_0_partial_block() {
    // Q8_0 with elements not divisible by block size (32)
    let mut data = build_header(1, 0);
    data.extend(build_tensor_info("test", &[50], GGUF_TYPE_Q8_0, 0)); // 50 = 1.56 blocks
    align_to_boundary(&mut data);

    // Need 2 blocks for 50 elements
    // Each Q8_0 block: 2 bytes f16 scale + 32 bytes quants = 34 bytes
    for _ in 0..2 {
        let scale = half::f16::from_f32(0.5);
        data.extend_from_slice(&scale.to_le_bytes());
        data.extend([0i8 as u8; 32]);
    }

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    let tensor = model.get_tensor_f32("test", &data).expect("Should extract");
    assert_eq!(tensor.len(), 50);
}

#[test]
fn test_loader_get_tensor_f16_values() {
    let mut data = build_header(1, 0);
    data.extend(build_tensor_info("test", &[8], GGUF_TYPE_F16, 0));
    align_to_boundary(&mut data);

    // 8 f16 values
    let values = [0.0f32, 0.5, 1.0, 1.5, 2.0, -1.0, -0.5, 0.25];
    for v in &values {
        let f16_val = half::f16::from_f32(*v);
        data.extend_from_slice(&f16_val.to_le_bytes());
    }

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    let tensor = model.get_tensor_f32("test", &data).expect("Should extract");
    assert_eq!(tensor.len(), 8);
    assert!((tensor[2] - 1.0).abs() < 0.01);
    assert!((tensor[5] - (-1.0)).abs() < 0.01);
}

#[test]
fn test_loader_get_tensor_q4_1_block() {
    let mut data = build_header(1, 0);
    data.extend(build_tensor_info("test", &[32], GGUF_TYPE_Q4_1, 0));
    align_to_boundary(&mut data);

    // Q4_1: 2 f16 scale + 2 f16 min + 16 bytes quants = 20 bytes
    let scale = half::f16::from_f32(1.0);
    let min = half::f16::from_f32(0.0);
    data.extend_from_slice(&scale.to_le_bytes());
    data.extend_from_slice(&min.to_le_bytes());
    data.extend([0x00u8; 16]);

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    let tensor = model.get_tensor_f32("test", &data).expect("Should extract");
    assert_eq!(tensor.len(), 32);
}

#[test]
fn test_loader_get_tensor_q5_0_block() {
    let mut data = build_header(1, 0);
    data.extend(build_tensor_info("test", &[32], GGUF_TYPE_Q5_0, 0));
    align_to_boundary(&mut data);

    // Q5_0: 2 f16 scale + 4 high bits + 16 bytes quants = 22 bytes
    let scale = half::f16::from_f32(1.0);
    data.extend_from_slice(&scale.to_le_bytes());
    data.extend([0u8; 4]); // high bits
    data.extend([0x00u8; 16]); // quants

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    let tensor = model.get_tensor_f32("test", &data).expect("Should extract");
    assert_eq!(tensor.len(), 32);
}

#[test]
fn test_loader_get_tensor_q5_1_block() {
    let mut data = build_header(1, 0);
    data.extend(build_tensor_info("test", &[32], GGUF_TYPE_Q5_1, 0));
    align_to_boundary(&mut data);

    // Q5_1: 2 f16 scale + 2 f16 min + 4 high bits + 16 bytes quants = 24 bytes
    let scale = half::f16::from_f32(1.0);
    let min = half::f16::from_f32(0.0);
    data.extend_from_slice(&scale.to_le_bytes());
    data.extend_from_slice(&min.to_le_bytes());
    data.extend([0u8; 4]); // high bits
    data.extend([0x00u8; 16]); // quants

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    let tensor = model.get_tensor_f32("test", &data).expect("Should extract");
    assert_eq!(tensor.len(), 32);
}

#[test]
fn test_loader_get_tensor_k_quant_partial() {
    // K-quant with partial super-block (elements not divisible by 256)
    let mut data = build_header(1, 0);
    data.extend(build_tensor_info("test", &[300], GGUF_TYPE_Q4_K, 0)); // 300 = 1.17 super-blocks
    align_to_boundary(&mut data);

    // Need 2 super-blocks (512 elements worth)
    // Q4_K super-block: 144 bytes
    data.extend([0u8; 288]); // 2 * 144

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    let tensor = model.get_tensor_f32("test", &data).expect("Should extract");
    assert_eq!(tensor.len(), 300); // Truncated
}

#[test]
fn test_loader_get_tensor_offset_nonzero() {
    // Tensor at non-zero offset
    let mut data = build_header(1, 0);
    // Offset 64 bytes into tensor data section
    data.extend(build_tensor_info("test", &[4], GGUF_TYPE_F32, 64));
    align_to_boundary(&mut data);

    // Padding before tensor data
    data.extend([0xFFu8; 64]);
    // Actual tensor data
    for i in 0..4 {
        data.extend_from_slice(&((i + 100) as f32).to_le_bytes());
    }

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    let tensor = model.get_tensor_f32("test", &data).expect("Should extract");
    assert_eq!(tensor.len(), 4);
    assert!((tensor[0] - 100.0).abs() < f32::EPSILON);
}

// =============================================================================
// Error Message Quality Tests
// =============================================================================

#[test]
fn test_loader_error_tensor_not_found_message() {
    let data = build_header(0, 0);
    let model = GGUFModel::from_bytes(&data).expect("Should parse");

    let result = model.get_tensor_f32("nonexistent_tensor", &data);
    assert!(result.is_err());

    let err = result.unwrap_err();
    let err_str = err.to_string();
    assert!(
        err_str.contains("nonexistent_tensor") || err_str.contains("not found"),
        "Error should mention tensor name: {}",
        err_str
    );
}

#[test]
fn test_loader_error_unsupported_qtype_message() {
    let mut data = build_header(1, 0);
    data.extend(build_tensor_info("test", &[32], 200, 0)); // qtype 200 doesn't exist
    align_to_boundary(&mut data);
    data.extend([0u8; 128]);

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    let result = model.get_tensor_f32("test", &data);

    assert!(result.is_err());
    let err_str = result.unwrap_err().to_string();
    assert!(
        err_str.contains("200") || err_str.contains("Unsupported"),
        "Error should mention unsupported type: {}",
        err_str
    );
}

#[test]
fn test_loader_error_data_range_message() {
    let mut data = build_header(1, 0);
    // Claim huge tensor that exceeds file size
    data.extend(build_tensor_info("huge", &[1_000_000], GGUF_TYPE_F32, 0));
    // No actual tensor data

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    let result = model.get_tensor_f32("huge", &data);

    assert!(result.is_err());
    let err_str = result.unwrap_err().to_string();
    assert!(
        err_str.contains("exceeds") || err_str.contains("range"),
        "Error should mention data range issue: {}",
        err_str
    );
}

// =============================================================================
// Edge Cases in String Parsing
// =============================================================================

#[test]
fn test_loader_string_with_null_bytes() {
    let mut data = build_header(0, 1);
    // String containing null bytes
    let s = "hello\0world";
    let str_bytes = build_string(s);
    data.extend(build_metadata("null_str", 8, &str_bytes));

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    if let Some(GGUFValue::String(val)) = model.metadata.get("null_str") {
        assert_eq!(val, "hello\0world");
    }
}

#[test]
fn test_loader_string_utf8_multibyte() {
    let mut data = build_header(0, 1);
    // Various UTF-8 multibyte sequences
    let s = "\u{00E9}\u{4E2D}\u{1F600}"; // e-acute, Chinese char, emoji
    let str_bytes = build_string(s);
    data.extend(build_metadata("utf8", 8, &str_bytes));

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    if let Some(GGUFValue::String(val)) = model.metadata.get("utf8") {
        assert_eq!(val, s);
    }
}

#[test]
fn test_loader_tensor_name_with_dots() {
    let mut data = build_header(1, 0);
    // Typical GGUF tensor name with multiple dots
    data.extend(build_tensor_info(
        "blk.15.attn_q.weight",
        &[4096, 4096],
        GGUF_TYPE_Q4_K,
        0,
    ));

    let model = GGUFModel::from_bytes(&data).expect("Should parse");
    assert_eq!(model.tensors[0].name, "blk.15.attn_q.weight");
}