realizar 0.8.5

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

//! Dequantization functions for GGUF quantization formats (PMAT-802)
//!
//! Extracted from quantize/mod.rs - Basic dequantization for Q4_0, Q8_0, F16,
//! Q4_1, Q5_0, Q5_1, Q4_K, Q5_K, Q6_K, Q2_K formats.
//!
//! ## Contents
//! - `dequantize_q4_0` - 4-bit quantization (block size 32)
//! - `dequantize_q8_0` - 8-bit quantization (block size 32)
//! - `dequantize_f16` - Half-precision to f32
//! - `dequantize_q4_1` - 4-bit with scale and min
//! - `dequantize_q5_0` - 5-bit quantization
//! - `dequantize_q5_1` - 5-bit with scale and min
//! - `dequantize_q4_k` - K-quantization 4-bit (super-block 256)
//! - `dequantize_q5_k` - K-quantization 5-bit (super-block 256)
//! - `dequantize_q6_k` - K-quantization 6-bit (super-block 256)
//! - `dequantize_q2_k` - K-quantization 2-bit (super-block 256)

use crate::error::{RealizarError, Result};
use crate::quantize::{BLOCK_SIZE, QK_K};

// Re-use helpers from simd module
use super::simd::extract_scale_min;

/// Dequantize `Q4_0` format weights
///
/// # Arguments
///
/// * `data` - Raw `Q4_0` quantized data (blocks of scale + 16 bytes)
///
/// # Returns
///
/// Dequantized float32 values
///
/// # Errors
///
/// Returns error if data length is not a multiple of block size
///
/// # Examples
///
/// ```rust,ignore
/// let quantized = load_q4_0_weights();
/// let weights = dequantize_q4_0(&quantized)?;
/// ```
pub fn dequantize_q4_0(data: &[u8]) -> Result<Vec<f32>> {
    // Q4_0 block: 2 bytes (f16 scale) + 16 bytes (quants) = 18 bytes
    // GGML spec: typedef struct { ggml_half d; uint8_t qs[QK4_0/2]; } block_q4_0;
    const BLOCK_BYTES: usize = 2 + 16;

    if !data.len().is_multiple_of(BLOCK_BYTES) {
        return Err(RealizarError::InvalidShape {
            reason: format!(
                "Q4_0 data length {} is not a multiple of block size {}",
                data.len(),
                BLOCK_BYTES
            ),
        });
    }

    let num_blocks = data.len() / BLOCK_BYTES;
    let mut result = vec![0.0f32; num_blocks * BLOCK_SIZE];

    for block_idx in 0..num_blocks {
        let block_start = block_idx * BLOCK_BYTES;
        let out_start = block_idx * BLOCK_SIZE;

        // Read scale (f16, per GGML spec)
        let scale_bytes = &data[block_start..block_start + 2];
        let scale = half::f16::from_le_bytes([scale_bytes[0], scale_bytes[1]]).to_f32();

        // Read quantized values (16 bytes)
        let quants_start = block_start + 2;
        let quants = &data[quants_start..quants_start + 16];

        // Dequantize following candle's layout:
        // - Positions 0-15: low nibbles of bytes 0-15
        // - Positions 16-31: high nibbles of bytes 0-15
        for (j, &byte) in quants.iter().enumerate() {
            // Low 4 bits go to position j
            #[allow(clippy::cast_possible_wrap)]
            let low = (byte & 0x0F) as i16 - 8;
            result[out_start + j] = scale * (low as f32);

            // High 4 bits go to position j + 16
            #[allow(clippy::cast_possible_wrap)]
            let high = (byte >> 4) as i16 - 8;
            result[out_start + j + 16] = scale * (high as f32);
        }
    }

    Ok(result)
}

/// Dequantize `Q8_0` format weights
///
/// # Arguments
///
/// * `data` - Raw `Q8_0` quantized data (blocks of scale + 32 int8 values)
///
/// # Returns
///
/// Dequantized float32 values
///
/// # Errors
///
/// Returns error if data length is not a multiple of block size
pub fn dequantize_q8_0(data: &[u8]) -> Result<Vec<f32>> {
    // Q8_0 block: 2 bytes (f16 scale) + 32 bytes (int8 quants) = 34 bytes
    // Note: GGML spec uses f16 for scale, not f32!
    const BLOCK_BYTES: usize = 2 + 32;

    if !data.len().is_multiple_of(BLOCK_BYTES) {
        return Err(RealizarError::InvalidShape {
            reason: format!(
                "Q8_0 data length {} is not a multiple of block size {}",
                data.len(),
                BLOCK_BYTES
            ),
        });
    }

    let num_blocks = data.len() / BLOCK_BYTES;
    let mut result = Vec::with_capacity(num_blocks * BLOCK_SIZE);

    for block_idx in 0..num_blocks {
        let block_start = block_idx * BLOCK_BYTES;

        // Read scale (f16 -> f32)
        let scale_bits = u16::from_le_bytes([data[block_start], data[block_start + 1]]);
        let scale = f16_to_f32(scale_bits);

        // Read quantized values (32 int8 values)
        let quants_start = block_start + 2;
        let quants = &data[quants_start..quants_start + 32];

        // Dequantize
        for &byte in quants {
            let value = i8::from_le_bytes([byte]);
            result.push(scale * f32::from(value));
        }
    }

    Ok(result)
}

/// Convert IEEE 754 half-precision (f16) to single-precision (f32)
///
/// ONE PATH: Delegates to `trueno::f16_to_f32` (UCBD §4).
#[inline]
pub fn f16_to_f32(h: u16) -> f32 {
    trueno::f16_to_f32(h)
}

/// Dequantize `F16` format weights to `F32`
///
/// # Arguments
///
/// * `data` - Raw F16 data (2 bytes per value)
///
/// # Returns
///
/// Dequantized float32 values
///
/// # Errors
///
/// Returns error if data length is not a multiple of 2 bytes
pub fn dequantize_f16(data: &[u8]) -> Result<Vec<f32>> {
    if !data.len().is_multiple_of(2) {
        return Err(RealizarError::InvalidShape {
            reason: format!(
                "F16 data length {} is not a multiple of 2 bytes",
                data.len()
            ),
        });
    }

    let num_values = data.len() / 2;
    let mut result = Vec::with_capacity(num_values);

    for chunk in data.chunks_exact(2) {
        let h = u16::from_le_bytes([chunk[0], chunk[1]]);
        result.push(f16_to_f32(h));
    }

    Ok(result)
}

/// Dequantize `Q4_1` format weights
///
/// Q4_1 format: 2 bytes (f16 scale) + 2 bytes (f16 min) + 16 bytes (quants) = 20 bytes
/// GGUF/candle layout:
/// - Positions 0-15: low nibbles of bytes 0-15
/// - Positions 16-31: high nibbles of bytes 0-15
pub fn dequantize_q4_1(data: &[u8]) -> Result<Vec<f32>> {
    const BLOCK_BYTES: usize = 20;

    if !data.len().is_multiple_of(BLOCK_BYTES) {
        return Err(RealizarError::InvalidShape {
            reason: format!(
                "Q4_1 data length {} is not a multiple of block size {}",
                data.len(),
                BLOCK_BYTES
            ),
        });
    }

    let num_blocks = data.len() / BLOCK_BYTES;
    // Pre-allocate with correct size for candle layout
    let mut result = vec![0.0f32; num_blocks * BLOCK_SIZE];

    for block_idx in 0..num_blocks {
        let block_start = block_idx * BLOCK_BYTES;
        let out_start = block_idx * BLOCK_SIZE;

        let d_bytes = &data[block_start..block_start + 2];
        let d = f16_to_f32(u16::from_le_bytes([d_bytes[0], d_bytes[1]]));

        let min_bytes = &data[block_start + 2..block_start + 4];
        let min = f16_to_f32(u16::from_le_bytes([min_bytes[0], min_bytes[1]]));

        let quants = &data[block_start + 4..block_start + 20];

        // Use candle layout (same as Q4_0):
        // - Low nibbles (byte & 0xF) at positions 0-15
        // - High nibbles (byte >> 4) at positions 16-31
        for (j, &byte) in quants.iter().enumerate() {
            // Low 4 bits go to position j (0-15)
            let low = byte & 0x0F;
            result[out_start + j] = d * f32::from(low) + min;

            // High 4 bits go to position j + 16 (16-31)
            let high = (byte >> 4) & 0x0F;
            result[out_start + j + 16] = d * f32::from(high) + min;
        }
    }

    Ok(result)
}

/// Dequantize `Q5_0` format weights
///
/// Q5_0 format: 2 bytes (f16 scale) + 4 bytes (high bits) + 16 bytes (quants) = 22 bytes
/// GGUF/candle layout:
/// - Positions 0-15: low nibbles + high bits from qh
/// - Positions 16-31: high nibbles + high bits from qh
pub fn dequantize_q5_0(data: &[u8]) -> Result<Vec<f32>> {
    const BLOCK_BYTES: usize = 22;

    if !data.len().is_multiple_of(BLOCK_BYTES) {
        return Err(RealizarError::InvalidShape {
            reason: format!(
                "Q5_0 data length {} is not a multiple of block size {}",
                data.len(),
                BLOCK_BYTES
            ),
        });
    }

    let num_blocks = data.len() / BLOCK_BYTES;
    // Pre-allocate with correct size for candle layout
    let mut result = vec![0.0f32; num_blocks * BLOCK_SIZE];

    for block_idx in 0..num_blocks {
        let block_start = block_idx * BLOCK_BYTES;
        let out_start = block_idx * BLOCK_SIZE;

        let d_bytes = &data[block_start..block_start + 2];
        let d = f16_to_f32(u16::from_le_bytes([d_bytes[0], d_bytes[1]]));

        let qh = u32::from_le_bytes([
            data[block_start + 2],
            data[block_start + 3],
            data[block_start + 4],
            data[block_start + 5],
        ]);

        let qs = &data[block_start + 6..block_start + 22];

        // Use candle layout:
        // - Low nibbles (byte & 0xF) at positions 0-15
        // - High nibbles (byte >> 4) at positions 16-31
        for (i, &byte) in qs.iter().enumerate() {
            // Low 4 bits + 5th bit go to position i (0-15)
            let low_q = byte & 0x0F;
            let high_bit_low = ((qh >> i) & 1) as u8;
            let q_low = low_q | (high_bit_low << 4);
            #[allow(clippy::cast_possible_wrap)]
            let value_low = q_low as i8 - 16;
            result[out_start + i] = d * f32::from(value_low);

            // High 4 bits + 5th bit go to position i + 16 (16-31)
            let high_q = (byte >> 4) & 0x0F;
            let high_bit_high = ((qh >> (i + 16)) & 1) as u8;
            let q_high = high_q | (high_bit_high << 4);
            #[allow(clippy::cast_possible_wrap)]
            let value_high = q_high as i8 - 16;
            result[out_start + i + 16] = d * f32::from(value_high);
        }
    }

    Ok(result)
}

/// Dequantize `Q5_1` format weights
///
/// Q5_1 format: 2 bytes (f16 scale) + 2 bytes (f16 min) + 4 bytes (high bits) + 16 bytes (quants) = 24 bytes
/// GGUF/candle layout:
/// - Positions 0-15: low nibbles + high bits from qh
/// - Positions 16-31: high nibbles + high bits from qh
pub fn dequantize_q5_1(data: &[u8]) -> Result<Vec<f32>> {
    const BLOCK_BYTES: usize = 24;

    if !data.len().is_multiple_of(BLOCK_BYTES) {
        return Err(RealizarError::InvalidShape {
            reason: format!(
                "Q5_1 data length {} is not a multiple of block size {}",
                data.len(),
                BLOCK_BYTES
            ),
        });
    }

    let num_blocks = data.len() / BLOCK_BYTES;
    // Pre-allocate with correct size for candle layout
    let mut result = vec![0.0f32; num_blocks * BLOCK_SIZE];

    for block_idx in 0..num_blocks {
        let block_start = block_idx * BLOCK_BYTES;
        let out_start = block_idx * BLOCK_SIZE;

        let d_bytes = &data[block_start..block_start + 2];
        let d = f16_to_f32(u16::from_le_bytes([d_bytes[0], d_bytes[1]]));

        let min_bytes = &data[block_start + 2..block_start + 4];
        let min = f16_to_f32(u16::from_le_bytes([min_bytes[0], min_bytes[1]]));

        let qh = u32::from_le_bytes([
            data[block_start + 4],
            data[block_start + 5],
            data[block_start + 6],
            data[block_start + 7],
        ]);

        let qs = &data[block_start + 8..block_start + 24];

        // Use candle layout:
        // - Low nibbles (byte & 0xF) at positions 0-15
        // - High nibbles (byte >> 4) at positions 16-31
        for (i, &byte) in qs.iter().enumerate() {
            // Low 4 bits + 5th bit go to position i (0-15)
            let low_q = byte & 0x0F;
            let high_bit_low = ((qh >> i) & 1) as u8;
            let q_low = low_q | (high_bit_low << 4);
            result[out_start + i] = d * f32::from(q_low) + min;

            // High 4 bits + 5th bit go to position i + 16 (16-31)
            let high_q = (byte >> 4) & 0x0F;
            let high_bit_high = ((qh >> (i + 16)) & 1) as u8;
            let q_high = high_q | (high_bit_high << 4);
            result[out_start + i + 16] = d * f32::from(q_high) + min;
        }
    }

    Ok(result)
}

include!("dequant_q4k.rs");
include!("dequant_f16.rs");