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
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
impl GGUFModel {
/// Extract tensor data by name with dequantization
///
/// # Arguments
///
/// * `name` - Tensor name to extract
/// * `file_data` - Complete GGUF file bytes
///
/// # Returns
///
/// Dequantized f32 tensor data
///
/// # Errors
///
/// Returns error if:
/// - Tensor not found
/// - Unsupported quantization type
/// - Invalid data at offset
///
/// # Examples
///
/// ```rust,ignore
/// let file_data = std::fs::read("model.gguf")?;
/// let model = GGUFModel::from_bytes(&file_data)?;
/// let weights = model.get_tensor_f32("layer.0.weight", &file_data)?;
/// ```
pub fn get_tensor_f32(&self, name: &str, file_data: &[u8]) -> Result<Vec<f32>> {
// Find tensor info
let tensor = self
.tensors
.iter()
.find(|t| t.name == name)
.ok_or_else(|| RealizarError::UnsupportedOperation {
operation: "get_tensor_f32".to_string(),
reason: format!("Tensor '{name}' not found"),
})?;
// Calculate tensor size in elements
let size: usize = tensor
.dims
.iter()
.try_fold(1usize, |acc, &dim| {
usize::try_from(dim).ok().and_then(|d| acc.checked_mul(d))
})
.ok_or_else(|| RealizarError::InvalidShape {
reason: format!("Tensor dimensions overflow: {:?}", tensor.dims),
})?;
// Convert tensor offset to usize and add tensor data start
let tensor_offset =
usize::try_from(tensor.offset).map_err(|_| RealizarError::UnsupportedOperation {
operation: "convert_offset".to_string(),
reason: format!("Offset {} exceeds platform usize limit", tensor.offset),
})?;
let offset = self.tensor_data_start + tensor_offset;
// Extract and dequantize based on qtype
match tensor.qtype {
GGUF_TYPE_F32 => {
// Unquantized F32 data
let byte_size = size * 4; // 4 bytes per f32
if offset + byte_size > file_data.len() {
return Err(RealizarError::UnsupportedOperation {
operation: "get_tensor_f32".to_string(),
reason: format!(
"Data range [{}, {}) exceeds file size {}",
offset,
offset + byte_size,
file_data.len()
),
});
}
let bytes = &file_data[offset..offset + byte_size];
let values = bytes
.chunks_exact(4)
.map(|chunk| f32::from_le_bytes([chunk[0], chunk[1], chunk[2], chunk[3]]))
.collect();
Ok(values)
},
GGUF_TYPE_Q4_0 => {
// Q4_0 quantized data
use crate::quantize::dequantize_q4_0;
// Q4_0 block: 32 elements
// Layout: 1×f16 scale (2 bytes) + 16 bytes (32×4-bit values) = 18 bytes
const BLOCK_BYTES: usize = 18;
const BLOCK_SIZE: usize = 32;
let num_blocks = size.div_ceil(BLOCK_SIZE);
let byte_size = num_blocks * BLOCK_BYTES;
if offset + byte_size > file_data.len() {
return Err(RealizarError::UnsupportedOperation {
operation: "get_tensor_f32".to_string(),
reason: format!(
"Data range [{}, {}) exceeds file size {}",
offset,
offset + byte_size,
file_data.len()
),
});
}
let bytes = &file_data[offset..offset + byte_size];
let mut values = dequantize_q4_0(bytes)?;
// Trim to exact size (dequantization pads to block boundaries)
values.truncate(size);
Ok(values)
},
GGUF_TYPE_Q8_0 => {
// Q8_0 quantized data - use SIMD-parallel for faster loading
use crate::quantize::dequantize_q8_0_simd;
// Q8_0 block size: 34 bytes (2 for f16 scale + 32 for quants)
const BLOCK_BYTES: usize = 34;
const BLOCK_SIZE: usize = 32;
let num_blocks = size.div_ceil(BLOCK_SIZE);
let byte_size = num_blocks * BLOCK_BYTES;
if offset + byte_size > file_data.len() {
return Err(RealizarError::UnsupportedOperation {
operation: "get_tensor_f32".to_string(),
reason: format!(
"Data range [{}, {}) exceeds file size {}",
offset,
offset + byte_size,
file_data.len()
),
});
}
let bytes = &file_data[offset..offset + byte_size];
let mut values = dequantize_q8_0_simd(bytes)?;
// Trim to exact size
values.truncate(size);
Ok(values)
},
GGUF_TYPE_Q2_K => {
// Q2_K quantized data (K-quantization) - 2 bits per weight
use crate::quantize::{dequantize_q2_k, QK_K};
// Q2_K super-block size: 84 bytes for 256 values
const SUPER_BLOCK_BYTES: usize = 84;
let num_super_blocks = size.div_ceil(QK_K);
let byte_size = num_super_blocks * SUPER_BLOCK_BYTES;
if offset + byte_size > file_data.len() {
return Err(RealizarError::UnsupportedOperation {
operation: "get_tensor_f32".to_string(),
reason: format!(
"Data range [{}, {}) exceeds file size {}",
offset,
offset + byte_size,
file_data.len()
),
});
}
let bytes = &file_data[offset..offset + byte_size];
let mut values = dequantize_q2_k(bytes)?;
// Trim to exact size
values.truncate(size);
Ok(values)
},
GGUF_TYPE_Q4_K => {
// Q4_K quantized data (K-quantization) - use SIMD-parallel for faster loading
use crate::quantize::{dequantize_q4_k_simd, QK_K};
// Q4_K super-block size: 144 bytes for 256 values
const SUPER_BLOCK_BYTES: usize = 144;
let num_super_blocks = size.div_ceil(QK_K);
let byte_size = num_super_blocks * SUPER_BLOCK_BYTES;
if offset + byte_size > file_data.len() {
return Err(RealizarError::UnsupportedOperation {
operation: "get_tensor_f32".to_string(),
reason: format!(
"Data range [{}, {}) exceeds file size {}",
offset,
offset + byte_size,
file_data.len()
),
});
}
let bytes = &file_data[offset..offset + byte_size];
let mut values = dequantize_q4_k_simd(bytes)?;
// Trim to exact size
values.truncate(size);
Ok(values)
},
GGUF_TYPE_Q5_K => {
// Q5_K quantized data (K-quantization)
use crate::quantize::{dequantize_q5_k, QK_K};
// Q5_K super-block size: 176 bytes for 256 values
const SUPER_BLOCK_BYTES: usize = 176;
let num_super_blocks = size.div_ceil(QK_K);
let byte_size = num_super_blocks * SUPER_BLOCK_BYTES;
if offset + byte_size > file_data.len() {
return Err(RealizarError::UnsupportedOperation {
operation: "get_tensor_f32".to_string(),
reason: format!(
"Data range [{}, {}) exceeds file size {}",
offset,
offset + byte_size,
file_data.len()
),
});
}
let bytes = &file_data[offset..offset + byte_size];
let mut values = dequantize_q5_k(bytes)?;
// Trim to exact size
values.truncate(size);
Ok(values)
},
GGUF_TYPE_Q6_K => {
// Q6_K quantized data (K-quantization)
use crate::quantize::{dequantize_q6_k, QK_K};
// Q6_K super-block size: 210 bytes for 256 values
const SUPER_BLOCK_BYTES: usize = 210;
let num_super_blocks = size.div_ceil(QK_K);
let byte_size = num_super_blocks * SUPER_BLOCK_BYTES;
if offset + byte_size > file_data.len() {
return Err(RealizarError::UnsupportedOperation {
operation: "get_tensor_f32".to_string(),
reason: format!(
"Data range [{}, {}) exceeds file size {}",
offset,
offset + byte_size,
file_data.len()
),
});
}
let bytes = &file_data[offset..offset + byte_size];
let mut values = dequantize_q6_k(bytes)?;
// Trim to exact size
values.truncate(size);
Ok(values)
},
GGUF_TYPE_F16 => {
// F16 (half-precision float) data
use crate::quantize::dequantize_f16;
let byte_size = size * 2; // 2 bytes per f16
if offset + byte_size > file_data.len() {
return Err(RealizarError::UnsupportedOperation {
operation: "get_tensor_f32".to_string(),
reason: format!(
"Data range [{}, {}) exceeds file size {}",
offset,
offset + byte_size,
file_data.len()
),
});
}
let bytes = &file_data[offset..offset + byte_size];
let values = dequantize_f16(bytes)?;
Ok(values)
},
GGUF_TYPE_Q4_1 => {
// Q4_1 quantized data
use crate::quantize::dequantize_q4_1;
// Q4_1 block size: 20 bytes (2 for scale + 2 for min + 16 for quants)
const BLOCK_BYTES: usize = 20;
const BLOCK_SIZE: usize = 32;
let num_blocks = size.div_ceil(BLOCK_SIZE);
let byte_size = num_blocks * BLOCK_BYTES;
if offset + byte_size > file_data.len() {
return Err(RealizarError::UnsupportedOperation {
operation: "get_tensor_f32".to_string(),
reason: format!(
"Data range [{}, {}) exceeds file size {}",
offset,
offset + byte_size,
file_data.len()
),
});
}
let bytes = &file_data[offset..offset + byte_size];
let mut values = dequantize_q4_1(bytes)?;
// Trim to exact size
values.truncate(size);
Ok(values)
},
GGUF_TYPE_Q5_0 => {
// Q5_0 quantized data
use crate::quantize::dequantize_q5_0;
// Q5_0 block size: 22 bytes (2 for scale + 4 for high bits + 16 for quants)
const BLOCK_BYTES: usize = 22;
const BLOCK_SIZE: usize = 32;
let num_blocks = size.div_ceil(BLOCK_SIZE);
let byte_size = num_blocks * BLOCK_BYTES;
if offset + byte_size > file_data.len() {
return Err(RealizarError::UnsupportedOperation {
operation: "get_tensor_f32".to_string(),
reason: format!(
"Data range [{}, {}) exceeds file size {}",
offset,
offset + byte_size,
file_data.len()
),
});
}
let bytes = &file_data[offset..offset + byte_size];
let mut values = dequantize_q5_0(bytes)?;
// Trim to exact size
values.truncate(size);
Ok(values)
},
GGUF_TYPE_Q5_1 => {
// Q5_1 quantized data
use crate::quantize::dequantize_q5_1;
// Q5_1 block size: 24 bytes (2 for scale + 2 for min + 4 for high bits + 16 for quants)
const BLOCK_BYTES: usize = 24;
const BLOCK_SIZE: usize = 32;
let num_blocks = size.div_ceil(BLOCK_SIZE);
let byte_size = num_blocks * BLOCK_BYTES;
if offset + byte_size > file_data.len() {
return Err(RealizarError::UnsupportedOperation {
operation: "get_tensor_f32".to_string(),
reason: format!(
"Data range [{}, {}) exceeds file size {}",
offset,
offset + byte_size,
file_data.len()
),
});
}
let bytes = &file_data[offset..offset + byte_size];
let mut values = dequantize_q5_1(bytes)?;
// Trim to exact size
values.truncate(size);
Ok(values)
},
_ => Err(RealizarError::UnsupportedOperation {
operation: "get_tensor_f32".to_string(),
reason: format!("Unsupported quantization type: {}", tensor.qtype),
}),
}
}
/// Extract model architecture from metadata
pub fn architecture(&self) -> Option<&str> {
if let Some(GGUFValue::String(arch)) = self.metadata.get(crate::gguf::keys::GENERAL_ARCHITECTURE) {
Some(arch.as_str())
} else {
None
}
}
/// Get embedding dimension from metadata
pub fn embedding_dim(&self) -> Option<usize> {
let arch = self.architecture()?;
let key = crate::gguf::keys::arch_key(arch, crate::gguf::keys::EMBEDDING_LENGTH);
if let Some(GGUFValue::UInt32(dim)) = self.metadata.get(&key) {
Some(*dim as usize)
} else {
None
}
}
/// Get number of layers from metadata
pub fn num_layers(&self) -> Option<usize> {
let arch = self.architecture()?;
let key = crate::gguf::keys::arch_key(arch, crate::gguf::keys::BLOCK_COUNT);
if let Some(GGUFValue::UInt32(count)) = self.metadata.get(&key) {
Some(*count as usize)
} else {
None
}
}
/// Get number of attention heads from metadata
pub fn num_heads(&self) -> Option<usize> {
let arch = self.architecture()?;
let key = crate::gguf::keys::arch_key(arch, crate::gguf::keys::ATTENTION_HEAD_COUNT);
if let Some(GGUFValue::UInt32(count)) = self.metadata.get(&key) {
Some(*count as usize)
} else {
None
}
}
/// Get context length from metadata
pub fn context_length(&self) -> Option<usize> {
let arch = self.architecture()?;
let key = crate::gguf::keys::arch_key(arch, crate::gguf::keys::CONTEXT_LENGTH);
if let Some(GGUFValue::UInt32(len)) = self.metadata.get(&key) {
Some(*len as usize)
} else {
None
}
}
/// Get number of key-value heads from metadata (for GQA)
pub fn num_kv_heads(&self) -> Option<usize> {
let arch = self.architecture()?;
let key = crate::gguf::keys::arch_key(arch, crate::gguf::keys::ATTENTION_HEAD_COUNT_KV);
if let Some(GGUFValue::UInt32(count)) = self.metadata.get(&key) {
Some(*count as usize)
} else {
None
}
}
/// Get attention key length (head dimension) from metadata.
///
/// This is the per-head dimension for Q/K projections. For most models
/// this equals `hidden_dim / num_heads`, but Qwen3-0.6B has `head_dim=128`
/// while `hidden_dim=1024` and `num_heads=16` (so `q_dim=2048 ≠ hidden_dim`).
///
/// GGUF key: `{arch}.attention.key_length`
pub fn key_length(&self) -> Option<usize> {
let arch = self.architecture()?;
let key = crate::gguf::keys::arch_key(arch, crate::gguf::keys::ATTENTION_KEY_LENGTH);
if let Some(GGUFValue::UInt32(len)) = self.metadata.get(&key) {
Some(*len as usize)
} else {
None
}
}
/// Get attention value length (value head dimension) from metadata.
///
/// GGUF key: `{arch}.attention.value_length`
pub fn value_length(&self) -> Option<usize> {
let arch = self.architecture()?;
let key = crate::gguf::keys::arch_key(arch, crate::gguf::keys::ATTENTION_VALUE_LENGTH);
if let Some(GGUFValue::UInt32(len)) = self.metadata.get(&key) {
Some(*len as usize)
} else {
None
}
}
/// Get RoPE frequency base from metadata
/// Different models use different bases (LLaMA: 10000, Qwen2: 1000000)
pub fn rope_freq_base(&self) -> Option<f32> {
let arch = self.architecture()?;
let key = crate::gguf::keys::arch_key(arch, crate::gguf::keys::ROPE_FREQ_BASE);
if let Some(GGUFValue::Float32(base)) = self.metadata.get(&key) {
Some(*base)
} else {
None
}
}
}