spec-stream-engine 0.1.1

Speculative decoding async execution engine, pipeline scheduler, and token streaming orchestrator
Documentation
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
//! # GGUF Quantization Engine
//!
//! Provides pure-Rust, zero-dependency de-quantization and high-performance dot product
//! kernels for standard GGUF quantization formats:
//! - **Q8_0**: 8-bit quantized weights with FP16 block scales (34 bytes per 32 elements).
//! - **Q4_0**: 4-bit quantized weights with FP16 block scales (18 bytes per 32 elements).

use crate::error::EngineError;
use crate::model::ModelEngine;
use spec_stream_core::{Logits, TokenId};

/// Number of elements per GGML quantization block.
pub const QK_K: usize = 32;

/// Converts an IEEE 754 half-precision (FP16) 16-bit integer to a single-precision (FP32) float.
#[inline]
pub fn f16_to_f32(h: u16) -> f32 {
    let sign = ((h >> 15) & 1) as u32;
    let exp = ((h >> 10) & 0x1F) as u32;
    let mant = (h & 0x3FF) as u32;

    let f_sign = sign << 31;
    if exp == 0 {
        if mant == 0 {
            return f32::from_bits(f_sign);
        }
        // Subnormal number
        let mut m = mant;
        let mut e = 0;
        while (m & 0x400) == 0 {
            m <<= 1;
            e += 1;
        }
        m &= 0x3FF;
        let f_exp = (127 - 15 + 1 - e) << 23;
        let f_mant = m << 13;
        return f32::from_bits(f_sign | f_exp | f_mant);
    } else if exp == 0x1F {
        let f_exp = 0xFF << 23;
        let f_mant = mant << 13;
        return f32::from_bits(f_sign | f_exp | f_mant);
    }

    let f_exp = (exp + 127 - 15) << 23;
    let f_mant = mant << 13;
    f32::from_bits(f_sign | f_exp | f_mant)
}

/// Converts a single-precision float (FP32) into an IEEE 754 half-precision (FP16) integer.
#[inline]
pub fn f32_to_f16(val: f32) -> u16 {
    let bits = val.to_bits();
    let sign = (bits >> 31) & 1;
    let exp = ((bits >> 23) & 0xFF) as i32;
    let mant = bits & 0x7FFFFF;

    if exp == 0xFF {
        // Inf or NaN
        let h_mant = if mant != 0 { 0x200 } else { 0 };
        return ((sign << 15) | (0x1F << 10) | h_mant) as u16;
    }

    let new_exp = exp - 127 + 15;
    if new_exp >= 31 {
        return ((sign << 15) | (0x1F << 10)) as u16;
    } else if new_exp <= 0 {
        if new_exp < -10 {
            return (sign << 15) as u16;
        }
        let m = (mant | 0x800000) >> (1 - new_exp);
        return ((sign << 15) | (m >> 13)) as u16;
    }

    ((sign << 15) | ((new_exp as u32) << 10) | (mant >> 13)) as u16
}

/// Quantization format type.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum QuantType {
    /// 8-bit quantization with FP16 scale per 32 elements.
    Q8_0,
    /// 4-bit quantization with FP16 scale per 32 elements.
    Q4_0,
}

/// Dequantizes a Q8_0 byte slice into an FP32 output buffer.
pub fn dequantize_q8_0(src: &[u8], dst: &mut [f32]) {
    assert_eq!(dst.len() % QK_K, 0, "Destination must be multiple of 32");
    let num_blocks = dst.len() / QK_K;
    assert_eq!(
        src.len(),
        num_blocks * 34,
        "Source length mismatch for Q8_0"
    );

    for b in 0..num_blocks {
        let src_offset = b * 34;
        let scale_bytes = [src[src_offset], src[src_offset + 1]];
        let scale = f16_to_f32(u16::from_le_bytes(scale_bytes));

        let dst_offset = b * QK_K;
        for i in 0..QK_K {
            let q = src[src_offset + 2 + i] as i8;
            dst[dst_offset + i] = q as f32 * scale;
        }
    }
}

/// Dequantizes a Q4_0 byte slice into an FP32 output buffer.
pub fn dequantize_q4_0(src: &[u8], dst: &mut [f32]) {
    assert_eq!(dst.len() % QK_K, 0, "Destination must be multiple of 32");
    let num_blocks = dst.len() / QK_K;
    assert_eq!(
        src.len(),
        num_blocks * 18,
        "Source length mismatch for Q4_0"
    );

    for b in 0..num_blocks {
        let src_offset = b * 18;
        let scale_bytes = [src[src_offset], src[src_offset + 1]];
        let scale = f16_to_f32(u16::from_le_bytes(scale_bytes));

        let dst_offset = b * QK_K;
        for i in 0..16 {
            let byte = src[src_offset + 2 + i];
            let q0 = (byte & 0x0F) as i8 - 8;
            let q1 = ((byte >> 4) & 0x0F) as i8 - 8;

            dst[dst_offset + i] = q0 as f32 * scale;
            dst[dst_offset + i + 16] = q1 as f32 * scale;
        }
    }
}

/// Computes the dot product between a Q8_0 quantized weight row and an FP32 vector directly without full dequantization.
pub fn vec_dot_q8_0(w_bytes: &[u8], x: &[f32]) -> f32 {
    #[cfg(target_arch = "x86_64")]
    {
        if is_x86_feature_detected!("avx2") && is_x86_feature_detected!("fma") {
            // SAFETY: Validated hardware feature presence via CPUID
            return unsafe { vec_dot_q8_0_avx2(w_bytes, x) };
        }
    }

    vec_dot_q8_0_scalar(w_bytes, x)
}

/// Fallback scalar implementation of Q8_0 dot product.
pub fn vec_dot_q8_0_scalar(w_bytes: &[u8], x: &[f32]) -> f32 {
    let num_blocks = x.len() / QK_K;
    let mut sum = 0.0f32;

    for b in 0..num_blocks {
        let src_offset = b * 34;
        let scale = f16_to_f32(u16::from_le_bytes([
            w_bytes[src_offset],
            w_bytes[src_offset + 1],
        ]));
        let x_offset = b * QK_K;

        let mut block_acc = 0.0f32;
        for i in 0..QK_K {
            let q = w_bytes[src_offset + 2 + i] as i8;
            block_acc += q as f32 * x[x_offset + i];
        }
        sum += block_acc * scale;
    }

    sum
}

#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2,fma")]
unsafe fn vec_dot_q8_0_avx2(w_bytes: &[u8], x: &[f32]) -> f32 {
    use std::arch::x86_64::*;

    let num_blocks = x.len() / QK_K;
    let mut sum = 0.0f32;

    for b in 0..num_blocks {
        let src_offset = b * 34;
        let scale = f16_to_f32(u16::from_le_bytes([
            w_bytes[src_offset],
            w_bytes[src_offset + 1],
        ]));
        let x_ptr = x.as_ptr().add(b * QK_K);
        let q_ptr = w_bytes.as_ptr().add(src_offset + 2);

        let mut acc = _mm256_setzero_ps();

        for i in 0..4 {
            let i8_slice = std::slice::from_raw_parts(q_ptr.add(i * 8), 8);
            let mut i64_buf = [0u8; 8];
            i64_buf.copy_from_slice(i8_slice);
            let q_i64 = i64::from_ne_bytes(i64_buf);
            let q128 = _mm_cvtsi64_si128(q_i64);

            let q256_i32 = _mm256_cvtepi8_epi32(q128);
            let q256_f32 = _mm256_cvtepi32_ps(q256_i32);
            let x256_f32 = _mm256_loadu_ps(x_ptr.add(i * 8));

            acc = _mm256_fmadd_ps(q256_f32, x256_f32, acc);
        }

        let h128 = _mm_add_ps(_mm256_castps256_ps128(acc), _mm256_extractf128_ps(acc, 1));
        let shuf = _mm_movehdup_ps(h128);
        let sums = _mm_add_ps(h128, shuf);
        let shuf2 = _mm_movehl_ps(sums, sums);
        let total = _mm_add_ss(sums, shuf2);
        let block_sum = _mm_cvtss_f32(total);

        sum += block_sum * scale;
    }

    sum
}

/// Computes the dot product between a Q4_0 quantized weight row and an FP32 vector directly without full dequantization.
pub fn vec_dot_q4_0(w_bytes: &[u8], x: &[f32]) -> f32 {
    let num_blocks = x.len() / QK_K;
    let mut sum = 0.0f32;

    for b in 0..num_blocks {
        let src_offset = b * 18;
        let scale = f16_to_f32(u16::from_le_bytes([
            w_bytes[src_offset],
            w_bytes[src_offset + 1],
        ]));
        let x_offset = b * QK_K;

        let mut block_acc = 0.0f32;
        for i in 0..16 {
            let byte = w_bytes[src_offset + 2 + i];
            let q0 = (byte & 0x0F) as i8 - 8;
            let q1 = ((byte >> 4) & 0x0F) as i8 - 8;

            block_acc += q0 as f32 * x[x_offset + i];
            block_acc += q1 as f32 * x[x_offset + i + 16];
        }
        sum += block_acc * scale;
    }

    sum
}

/// Quantizes an FP32 slice of 32 elements into a Q8_0 block (34 bytes).
pub fn quantize_q8_0_block(src: &[f32], dst: &mut [u8]) {
    assert_eq!(src.len(), QK_K);
    assert_eq!(dst.len(), 34);

    let mut amax = 0.0f32;
    for &val in src {
        amax = amax.max(val.abs());
    }

    let d = amax / 127.0;
    let id = if d != 0.0 { 1.0 / d } else { 0.0 };

    let f16_scale = f32_to_f16(d);
    let scale_bytes = f16_scale.to_le_bytes();
    dst[0] = scale_bytes[0];
    dst[1] = scale_bytes[1];

    for i in 0..QK_K {
        let q = (src[i] * id).round().clamp(-128.0, 127.0) as i8;
        dst[2 + i] = q as u8;
    }
}

/// A memory-efficient quantized linear projection model (Q8_0 or Q4_0).
/// Reduces LM head weight memory by up to 75%–87% compared to FP32.
#[derive(Debug, Clone)]
pub struct QuantizedLinearModel {
    pub name: String,
    pub vocab_size: usize,
    pub hidden_dim: usize,
    pub quant_type: QuantType,
    /// Raw quantized bytes matrix of shape [vocab_size, bytes_per_row].
    pub quantized_weights: Vec<u8>,
    pub bytes_per_row: usize,
    /// Unquantized or dequantized token embedding table.
    pub embeddings: Vec<f32>,
}

impl QuantizedLinearModel {
    /// Creates a Q8_0 quantized linear model from FP32 weights.
    pub fn from_f32_weights(
        name: impl Into<String>,
        vocab_size: usize,
        hidden_dim: usize,
        weights: &[f32],
        embeddings: Vec<f32>,
    ) -> Result<Self, EngineError> {
        if hidden_dim % QK_K != 0 {
            return Err(EngineError::InvalidConfig(format!(
                "Hidden dimension {} must be a multiple of {}",
                hidden_dim, QK_K
            )));
        }

        let num_blocks = hidden_dim / QK_K;
        let bytes_per_row = num_blocks * 34; // Q8_0
        let mut quantized_weights = vec![0u8; vocab_size * bytes_per_row];

        for row in 0..vocab_size {
            let row_offset_f32 = row * hidden_dim;
            let row_offset_bytes = row * bytes_per_row;

            for b in 0..num_blocks {
                let src_slice =
                    &weights[row_offset_f32 + b * QK_K..row_offset_f32 + (b + 1) * QK_K];
                let dst_slice = &mut quantized_weights
                    [row_offset_bytes + b * 34..row_offset_bytes + (b + 1) * 34];
                quantize_q8_0_block(src_slice, dst_slice);
            }
        }

        Ok(Self {
            name: name.into(),
            vocab_size,
            hidden_dim,
            quant_type: QuantType::Q8_0,
            quantized_weights,
            bytes_per_row,
            embeddings,
        })
    }

    /// Projects hidden states into logits using quantized dot products.
    pub fn project_hidden(&self, hidden: &[f32]) -> Logits {
        assert_eq!(hidden.len(), self.hidden_dim);
        let mut logits = Vec::with_capacity(self.vocab_size);

        for row in 0..self.vocab_size {
            let offset = row * self.bytes_per_row;
            let row_bytes = &self.quantized_weights[offset..offset + self.bytes_per_row];

            let dot = match self.quant_type {
                QuantType::Q8_0 => vec_dot_q8_0(row_bytes, hidden),
                QuantType::Q4_0 => vec_dot_q4_0(row_bytes, hidden),
            };
            logits.push(dot);
        }

        Logits::new(logits)
    }

    /// Computes logits directly for an input token ID.
    pub fn project_token(&self, token: TokenId) -> Logits {
        let token_idx = (token as usize) % self.vocab_size;
        let emb_offset = token_idx * self.hidden_dim;
        let hidden = &self.embeddings[emb_offset..emb_offset + self.hidden_dim];
        self.project_hidden(hidden)
    }
}

impl ModelEngine for QuantizedLinearModel {
    async fn forward_batch(&mut self, tokens: &[TokenId]) -> Result<Vec<Logits>, EngineError> {
        let mut out = Vec::with_capacity(tokens.len());
        for &tok in tokens {
            out.push(self.project_token(tok));
        }
        Ok(out)
    }

    async fn forward_token(&mut self, token: TokenId) -> Result<Logits, EngineError> {
        Ok(self.project_token(token))
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_f16_to_f32_roundtrip() {
        let test_vals = [0.0f32, 1.0, -1.0, 0.5, 2.5, -15.25, 100.0];
        for &v in &test_vals {
            let h = f32_to_f16(v);
            let back = f16_to_f32(h);
            assert!((v - back).abs() < 1e-2, "Failed for {v}: got {back}");
        }
    }

    #[test]
    fn test_q8_0_quantize_dequantize() {
        let mut original = [0.0f32; 32];
        for (i, val) in original.iter_mut().enumerate() {
            *val = (i as f32 - 16.0) * 0.25;
        }

        let mut block = [0u8; 34];
        quantize_q8_0_block(&original, &mut block);

        let mut dequantized = [0.0f32; 32];
        dequantize_q8_0(&block, &mut dequantized);

        for i in 0..32 {
            assert!(
                (original[i] - dequantized[i]).abs() < 0.05,
                "Diff too large at {i}: orig {}, got {}",
                original[i],
                dequantized[i]
            );
        }
    }

    #[test]
    fn test_quantized_linear_model() {
        let vocab_size = 16;
        let hidden_dim = 32;
        let size = vocab_size * hidden_dim;

        let weights: Vec<f32> = (0..size).map(|i| (i as f32 % 10.0) * 0.1).collect();
        let embeddings: Vec<f32> = (0..size).map(|i| ((i + 5) as f32 % 10.0) * 0.1).collect();

        let model = QuantizedLinearModel::from_f32_weights(
            "test-q8", vocab_size, hidden_dim, &weights, embeddings,
        )
        .expect("Failed to build quantized model");

        assert_eq!(model.bytes_per_row, 34);
        assert_eq!(model.quantized_weights.len(), 16 * 34);

        let logits = model.project_token(2);
        assert_eq!(logits.vocab_size(), 16);
    }

    #[test]
    fn test_vec_dot_q8_0_simd_equivalence() {
        let mut original = [0.0f32; 64];
        let mut x = [0.0f32; 64];
        for i in 0..64 {
            original[i] = (i as f32 - 32.0) * 0.15;
            x[i] = (i as f32 % 7.0) * 0.3;
        }

        let mut w_bytes = vec![0u8; 68]; // 2 blocks of 34 bytes
        quantize_q8_0_block(&original[0..32], &mut w_bytes[0..34]);
        quantize_q8_0_block(&original[32..64], &mut w_bytes[34..68]);

        let scalar_res = vec_dot_q8_0_scalar(&w_bytes, &x);
        let simd_res = vec_dot_q8_0(&w_bytes, &x);

        assert!(
            (scalar_res - simd_res).abs() < 1e-4,
            "SIMD dot product ({simd_res}) must match scalar dot product ({scalar_res})"
        );
    }
}