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
/// Dequantize `Q6_K` format (K-quants)
/// `Q6_K`: super blocks of 256 elements
/// Each super block: ql (128 bytes) + qh (64 bytes) + scales (16 bytes) + d (f16) = 210 bytes
///
/// Delegates to `trueno_quant::dequantize_q6_k_to_f32` — the single source of truth.
#[ensures(ret.as_ref().map_or(true, |v| v.len() == num_elements))]
pub(crate) fn dequantize_q6_k(data: &[u8], start: usize, num_elements: usize) -> Result<Vec<f32>> {
const SUPER_BLOCK_SIZE: usize = 256;
const SUPER_BLOCK_BYTES: usize = 210;
let num_blocks = num_elements.div_ceil(SUPER_BLOCK_SIZE);
let total_bytes = num_blocks * SUPER_BLOCK_BYTES;
if start + total_bytes > data.len() {
return Err(AprenderError::FormatError {
message: "Q6_K data exceeds file size".to_string(),
});
}
Ok(trueno_quant::dequantize_q6_k_to_f32(
&data[start..],
num_elements,
))
}
/// Dequantize `Q4_1` format
/// `Q4_1`: blocks of 32 elements, each block has f16 scale + f16 min + 16 bytes of 4-bit quants
///
/// PMAT-231 FIX: Element order matches llama.cpp/GGML layout:
/// - Low nibbles first (elements 0-15)
/// - High nibbles second (elements 16-31)
#[ensures(ret.as_ref().map_or(true, |v| v.len() == num_elements))]
pub fn dequantize_q4_1(data: &[u8], start: usize, num_elements: usize) -> Result<Vec<f32>> {
const BLOCK_SIZE: usize = 32;
const BLOCK_BYTES: usize = 2 + 2 + 16; // f16 scale + f16 min + 16 bytes
let num_blocks = num_elements.div_ceil(BLOCK_SIZE);
let total_bytes = num_blocks * BLOCK_BYTES;
if start + total_bytes > data.len() {
return Err(AprenderError::FormatError {
message: "Q4_1 data exceeds file size".to_string(),
});
}
let mut result = Vec::with_capacity(num_elements);
let mut offset = start;
for _ in 0..num_blocks {
// GH-186 FIX: Use safe_f16_scale to clamp NaN/Inf/subnormal
let scale = safe_f16_scale(u16::from_le_bytes([data[offset], data[offset + 1]]));
let min = safe_f16_scale(u16::from_le_bytes([data[offset + 2], data[offset + 3]]));
offset += 4;
// PMAT-231: Low nibbles first (elements 0-15)
for i in 0..16 {
let byte = data[offset + i];
let v0 = f32::from(byte & 0x0F) * scale + min;
result.push(v0);
}
// PMAT-231: High nibbles second (elements 16-31)
for i in 0..16 {
let byte = data[offset + i];
let v1 = f32::from(byte >> 4) * scale + min;
result.push(v1);
}
offset += 16;
}
result.truncate(num_elements);
Ok(result)
}
/// Dequantize `Q2_K` format (K-quants)
/// `Q2_K`: super blocks of 256 elements
#[ensures(ret.as_ref().map_or(true, |v| v.len() == num_elements))]
pub(crate) fn dequantize_q2_k(data: &[u8], start: usize, num_elements: usize) -> Result<Vec<f32>> {
const SUPER_BLOCK_SIZE: usize = 256;
const SUPER_BLOCK_BYTES: usize = 2 + 2 + 16 + 64; // d, dmin, scales, qs = 84 bytes
let num_blocks = num_elements.div_ceil(SUPER_BLOCK_SIZE);
let total_bytes = num_blocks * SUPER_BLOCK_BYTES;
if start + total_bytes > data.len() {
return Err(AprenderError::FormatError {
message: "Q2_K data exceeds file size".to_string(),
});
}
let mut result = Vec::with_capacity(num_elements);
let mut offset = start;
for _ in 0..num_blocks {
// Read scales (16 bytes = 16 4-bit scale/min pairs)
let scales_bytes = &data[offset..offset + 16];
offset += 16;
// Read qs (64 bytes = 256 2-bit values)
let qs = &data[offset..offset + 64];
offset += 64;
// Read d and dmin
// GH-186 FIX: Use safe_f16_scale to clamp NaN/Inf/subnormal
let d = safe_f16_scale(u16::from_le_bytes([data[offset], data[offset + 1]]));
let dmin = safe_f16_scale(u16::from_le_bytes([data[offset + 2], data[offset + 3]]));
offset += 4;
// ggml `dequantize_row_q2_K` ordering (mirrors candle BlockQ2K::to_float
// and llama.cpp ggml-quants.c): two groups of 128 elements, each over a
// 32-byte qs window. Within a group, 4 sub-iterations at shift 0/2/4/6,
// each consuming TWO scale bytes — one for the window's low 16 bytes and
// one for its high 16 bytes. The previous "16 sub-blocks reading
// qs[j*4+l]" scheme applied the wrong scale to the wrong 2-bit lanes and
// produced corrupt output (185/256 elements wrong vs ggml).
let mut is = 0usize;
for group in 0..2 {
let chunk = &qs[group * 32..group * 32 + 32];
let mut shift = 0u8;
for _ in 0..4 {
let sc = scales_bytes[is];
is += 1;
let dl = d * f32::from(sc & 0x0F);
let ml = dmin * f32::from(sc >> 4);
for &q in &chunk[0..16] {
result.push(dl * f32::from((q >> shift) & 0x03) - ml);
}
let sc = scales_bytes[is];
is += 1;
let dl = d * f32::from(sc & 0x0F);
let ml = dmin * f32::from(sc >> 4);
for &q in &chunk[16..32] {
result.push(dl * f32::from((q >> shift) & 0x03) - ml);
}
shift += 2;
}
}
}
result.truncate(num_elements);
Ok(result)
}
/// Dequantize `Q3_K` format (K-quants)
/// `Q3_K`: super blocks of 256 elements
#[ensures(ret.as_ref().map_or(true, |v| v.len() == num_elements))]
pub(crate) fn dequantize_q3_k(data: &[u8], start: usize, num_elements: usize) -> Result<Vec<f32>> {
const SUPER_BLOCK_SIZE: usize = 256;
const SUPER_BLOCK_BYTES: usize = 32 + 64 + 12 + 2; // hmask, qs, scales, d = 110 bytes
let num_blocks = num_elements.div_ceil(SUPER_BLOCK_SIZE);
let total_bytes = num_blocks * SUPER_BLOCK_BYTES;
if start + total_bytes > data.len() {
return Err(AprenderError::FormatError {
message: "Q3_K data exceeds file size".to_string(),
});
}
let mut result = Vec::with_capacity(num_elements);
let mut offset = start;
for _ in 0..num_blocks {
// Read hmask (32 bytes = 256 high bits)
let hmask = &data[offset..offset + 32];
offset += 32;
// Read qs (64 bytes = 256 low 2-bit values)
let qs = &data[offset..offset + 64];
offset += 64;
// Read scales (12 bytes = packed 6-bit scales)
let scales_bytes = &data[offset..offset + 12];
offset += 12;
// Read d
// GH-186 FIX: Use safe_f16_scale to clamp NaN/Inf/subnormal
let d = safe_f16_scale(u16::from_le_bytes([data[offset], data[offset + 1]]));
offset += 2;
// Reconstruct the 16 SIX-bit signed scales from the 12 packed bytes via the GGML
// aux[] shuffle (kmask1/kmask2). The previous code read only the low/high nibbles of
// the first 8 bytes — 16 FOUR-bit values with a -8 offset — clipping every scale from
// [-32,31] to [-8,7] and dropping the upper 2 bits, so ~252/256 elements were wrong.
const KMASK1: u32 = 0x0303_0303;
const KMASK2: u32 = 0x0f0f_0f0f;
let mut aux = [
u32::from_le_bytes([scales_bytes[0], scales_bytes[1], scales_bytes[2], scales_bytes[3]]),
u32::from_le_bytes([scales_bytes[4], scales_bytes[5], scales_bytes[6], scales_bytes[7]]),
u32::from_le_bytes([
scales_bytes[8],
scales_bytes[9],
scales_bytes[10],
scales_bytes[11],
]),
0u32,
];
let tmp = aux[2];
aux[2] = ((aux[0] >> 4) & KMASK2) | (((tmp >> 4) & KMASK1) << 4);
aux[3] = ((aux[1] >> 4) & KMASK2) | (((tmp >> 6) & KMASK1) << 4);
aux[0] = (aux[0] & KMASK2) | ((tmp & KMASK1) << 4);
aux[1] = (aux[1] & KMASK2) | (((tmp >> 2) & KMASK1) << 4);
let mut scales = [0i8; 16];
for (w, word) in aux.iter().enumerate() {
for (k, &b) in word.to_le_bytes().iter().enumerate() {
scales[w * 4 + k] = b as i8; // 6-bit value; the -32 offset is applied below
}
}
// Dequantize: two 128-element halves (qs advances 32 bytes per half); within a half,
// four 32-element shift-blocks (shift = 0,2,4,6) each carrying two 16-lane scale
// groups; the high-mask bit `m` advances per shift-block. value = d*(scale-32)*(low-high)
// where high = 4 when the hmask bit is CLEAR, else 0.
let mut block_out = [0.0f32; 256];
let mut m: u32 = 1;
let mut is = 0usize;
for half in 0..2 {
let qs_half = &qs[half * 32..half * 32 + 32];
let out_half = half * 128;
let mut shift: u32 = 0;
for blk in 0..4 {
let out_blk = out_half + blk * 32;
for scale_index in 0..2 {
let dl = d * (f32::from(scales[is]) - 32.0);
let out_grp = out_blk + scale_index * 16;
for i in 0..16 {
let idx = i + 16 * scale_index;
let low = ((qs_half[idx] >> shift) & 3) as i8;
let high = if u32::from(hmask[idx]) & m == 0 { 4i8 } else { 0i8 };
block_out[out_grp + i] = dl * f32::from(low - high);
}
is += 1;
}
shift += 2;
m <<= 1;
}
}
result.extend_from_slice(&block_out);
}
result.truncate(num_elements);
Ok(result)
}
/// Approximate dequantization for I-quants (IQ2, IQ3, IQ4)
/// These use importance-weighted quantization with lookup tables.
/// For import purposes, we approximate with a simple linear mapping.
pub(crate) fn dequantize_iq_approximate(
data: &[u8],
start: usize,
num_elements: usize,
dtype: u32,
) -> Vec<f32> {
// I-quants have variable block sizes and complex lookup tables
// Approximate by treating as low-bit quantization with estimated scale
let (bits_per_element, block_size): (usize, usize) = match dtype {
13..=15 => (2, 256), // IQ2_XXS, IQ2_XS, IQ2_S
16 | 17 => (3, 256), // IQ3_XXS, IQ3_S
18 => (1, 256), // IQ1_S
_ => (4, 256), // IQ4_NL, IQ4_XS, and default
};
let bytes_per_block = (block_size * bits_per_element).div_ceil(8) + 4; // data + scale overhead
let num_blocks = num_elements.div_ceil(block_size);
// For approximation, create small random-ish values based on byte patterns
// This is NOT correct dequantization but allows import to proceed
let mut result = Vec::with_capacity(num_elements);
let scale = 0.01; // Small scale for approximate values
for block_idx in 0..num_blocks {
let block_start = start + block_idx * bytes_per_block;
for i in 0..block_size {
if result.len() >= num_elements {
break;
}
// Use byte pattern to generate approximate value
let byte_idx = block_start + (i * bits_per_element) / 8;
if byte_idx < data.len() {
let byte_val = data[byte_idx];
// Map to roughly centered distribution
let approx = (f32::from(byte_val) - 128.0) * scale;
result.push(approx);
} else {
result.push(0.0);
}
}
}
result.truncate(num_elements);
result
}