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
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
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
//! # Model Weights & Metadata Parsers
//!
//! Provides zero-copy / streaming parsers for GGUF (v2/v3) binary headers and
//! SafeTensors metadata, alongside lightweight weight loaders for neural model runtimes.

use crate::error::EngineError;
use crate::model::ModelEngine;
use spec_stream_core::{Logits, TokenId};
use std::collections::HashMap;
use std::io::{Read, Seek};

/// Magic identifier for GGUF files (`b"GGUF"`).
pub const GGUF_MAGIC: [u8; 4] = *b"GGUF";

/// Value type in GGUF metadata key-value pairs.
#[derive(Debug, Clone, PartialEq)]
pub enum GgufValue {
    Uint8(u8),
    Int8(i8),
    Uint16(u16),
    Int16(i16),
    Uint32(u32),
    Int32(i32),
    Float32(f32),
    Bool(bool),
    String(String),
    Array(Vec<GgufValue>),
    Uint64(u64),
    Int64(i64),
    Float64(f64),
}

/// Metadata describing a single tensor stored inside a GGUF file.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct GgufTensorInfo {
    pub name: String,
    pub dimensions: Vec<u64>,
    pub tensor_type: u32,
    pub offset: u64,
}

/// Parsed header and metadata of a GGUF model file.
#[derive(Debug, Clone)]
pub struct GgufMetadata {
    pub version: u32,
    pub tensor_count: u64,
    pub metadata_kv_count: u64,
    pub metadata: HashMap<String, GgufValue>,
    pub tensors: Vec<GgufTensorInfo>,
    pub data_offset: u64,
}

impl GgufMetadata {
    /// Parses a GGUF header from any reader with seek capabilities.
    pub fn parse<R: Read + Seek>(reader: &mut R) -> Result<Self, EngineError> {
        let mut magic = [0u8; 4];
        reader
            .read_exact(&mut magic)
            .map_err(|e| EngineError::Inference(format!("Failed to read GGUF magic: {}", e)))?;

        if magic != GGUF_MAGIC {
            return Err(EngineError::Inference(format!(
                "Invalid GGUF magic: expected {:?}, got {:?}",
                GGUF_MAGIC, magic
            )));
        }

        let mut buf4 = [0u8; 4];
        let mut buf8 = [0u8; 8];

        reader
            .read_exact(&mut buf4)
            .map_err(|e| EngineError::Inference(e.to_string()))?;
        let version = u32::from_le_bytes(buf4);
        if !(2..=3).contains(&version) {
            return Err(EngineError::Inference(format!(
                "Unsupported GGUF version: {}",
                version
            )));
        }

        reader
            .read_exact(&mut buf8)
            .map_err(|e| EngineError::Inference(e.to_string()))?;
        let tensor_count = u64::from_le_bytes(buf8);

        reader
            .read_exact(&mut buf8)
            .map_err(|e| EngineError::Inference(e.to_string()))?;
        let metadata_kv_count = u64::from_le_bytes(buf8);

        let mut metadata = HashMap::with_capacity(metadata_kv_count as usize);

        for _ in 0..metadata_kv_count {
            let key = read_gguf_string(reader)?;
            let val = read_gguf_value(reader)?;
            metadata.insert(key, val);
        }

        let mut tensors = Vec::with_capacity(tensor_count as usize);
        for _ in 0..tensor_count {
            let name = read_gguf_string(reader)?;
            reader
                .read_exact(&mut buf4)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            let n_dims = u32::from_le_bytes(buf4);

            let mut dims = Vec::with_capacity(n_dims as usize);
            for _ in 0..n_dims {
                reader
                    .read_exact(&mut buf8)
                    .map_err(|e| EngineError::Inference(e.to_string()))?;
                dims.push(u64::from_le_bytes(buf8));
            }

            reader
                .read_exact(&mut buf4)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            let tensor_type = u32::from_le_bytes(buf4);

            reader
                .read_exact(&mut buf8)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            let offset = u64::from_le_bytes(buf8);

            tensors.push(GgufTensorInfo {
                name,
                dimensions: dims,
                tensor_type,
                offset,
            });
        }

        let pos = reader
            .stream_position()
            .map_err(|e| EngineError::Inference(e.to_string()))?;
        let alignment = match metadata.get("general.alignment") {
            Some(GgufValue::Uint32(a)) => *a as u64,
            Some(GgufValue::Uint64(a)) => *a,
            _ => 32,
        };
        let remainder = pos % alignment;
        let data_offset = if remainder == 0 {
            pos
        } else {
            pos + (alignment - remainder)
        };

        Ok(Self {
            version,
            tensor_count,
            metadata_kv_count,
            metadata,
            tensors,
            data_offset,
        })
    }

    /// Retrieves an optional string parameter from GGUF metadata.
    pub fn get_string(&self, key: &str) -> Option<&str> {
        match self.metadata.get(key) {
            Some(GgufValue::String(s)) => Some(s.as_str()),
            _ => None,
        }
    }

    /// Retrieves an optional u64 parameter from GGUF metadata.
    pub fn get_u64(&self, key: &str) -> Option<u64> {
        match self.metadata.get(key) {
            Some(GgufValue::Uint64(v)) => Some(*v),
            Some(GgufValue::Uint32(v)) => Some(*v as u64),
            Some(GgufValue::Int64(v)) if *v >= 0 => Some(*v as u64),
            _ => None,
        }
    }
}

fn read_gguf_string<R: Read>(reader: &mut R) -> Result<String, EngineError> {
    let mut buf8 = [0u8; 8];
    reader
        .read_exact(&mut buf8)
        .map_err(|e| EngineError::Inference(e.to_string()))?;
    let len = u64::from_le_bytes(buf8) as usize;

    let mut str_bytes = vec![0u8; len];
    reader
        .read_exact(&mut str_bytes)
        .map_err(|e| EngineError::Inference(e.to_string()))?;

    String::from_utf8(str_bytes)
        .map_err(|e| EngineError::Inference(format!("GGUF invalid UTF-8 string: {}", e)))
}

fn read_gguf_value<R: Read>(reader: &mut R) -> Result<GgufValue, EngineError> {
    let mut buf4 = [0u8; 4];
    reader
        .read_exact(&mut buf4)
        .map_err(|e| EngineError::Inference(e.to_string()))?;
    let val_type = u32::from_le_bytes(buf4);

    read_gguf_value_typed(reader, val_type)
}

fn read_gguf_value_typed<R: Read>(reader: &mut R, val_type: u32) -> Result<GgufValue, EngineError> {
    let mut buf1 = [0u8; 1];
    let mut buf2 = [0u8; 2];
    let mut buf4 = [0u8; 4];
    let mut buf8 = [0u8; 8];

    match val_type {
        0 => {
            reader
                .read_exact(&mut buf1)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            Ok(GgufValue::Uint8(buf1[0]))
        }
        1 => {
            reader
                .read_exact(&mut buf1)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            Ok(GgufValue::Int8(buf1[0] as i8))
        }
        2 => {
            reader
                .read_exact(&mut buf2)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            Ok(GgufValue::Uint16(u16::from_le_bytes(buf2)))
        }
        3 => {
            reader
                .read_exact(&mut buf2)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            Ok(GgufValue::Int16(i16::from_le_bytes(buf2)))
        }
        4 => {
            reader
                .read_exact(&mut buf4)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            Ok(GgufValue::Uint32(u32::from_le_bytes(buf4)))
        }
        5 => {
            reader
                .read_exact(&mut buf4)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            Ok(GgufValue::Int32(i32::from_le_bytes(buf4)))
        }
        6 => {
            reader
                .read_exact(&mut buf4)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            Ok(GgufValue::Float32(f32::from_le_bytes(buf4)))
        }
        7 => {
            reader
                .read_exact(&mut buf1)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            Ok(GgufValue::Bool(buf1[0] != 0))
        }
        8 => {
            let s = read_gguf_string(reader)?;
            Ok(GgufValue::String(s))
        }
        9 => {
            reader
                .read_exact(&mut buf4)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            let elem_type = u32::from_le_bytes(buf4);
            reader
                .read_exact(&mut buf8)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            let count = u64::from_le_bytes(buf8) as usize;

            let mut arr = Vec::with_capacity(count);
            for _ in 0..count {
                arr.push(read_gguf_value_typed(reader, elem_type)?);
            }
            Ok(GgufValue::Array(arr))
        }
        10 => {
            reader
                .read_exact(&mut buf8)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            Ok(GgufValue::Uint64(u64::from_le_bytes(buf8)))
        }
        11 => {
            reader
                .read_exact(&mut buf8)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            Ok(GgufValue::Int64(i64::from_le_bytes(buf8)))
        }
        12 => {
            reader
                .read_exact(&mut buf8)
                .map_err(|e| EngineError::Inference(e.to_string()))?;
            Ok(GgufValue::Float64(f64::from_le_bytes(buf8)))
        }
        other => Err(EngineError::Inference(format!(
            "Unsupported GGUF value type ID: {}",
            other
        ))),
    }
}

/// Metadata description for SafeTensors files.
#[derive(Debug, Clone, serde::Deserialize, serde::Serialize)]
pub struct SafeTensorEntry {
    pub dtype: String,
    pub shape: Vec<usize>,
    pub data_offsets: [usize; 2],
}

/// Parsed SafeTensors header.
#[derive(Debug, Clone)]
pub struct SafeTensorsMetadata {
    pub tensors: HashMap<String, SafeTensorEntry>,
    pub data_offset: usize,
}

impl SafeTensorsMetadata {
    /// Parses SafeTensors header from bytes.
    pub fn parse(bytes: &[u8]) -> Result<Self, EngineError> {
        if bytes.len() < 8 {
            return Err(EngineError::Inference(
                "SafeTensors buffer too short for 8-byte header length".into(),
            ));
        }

        let header_len = u64::from_le_bytes(bytes[0..8].try_into().unwrap()) as usize;
        let data_offset = 8 + header_len;

        if bytes.len() < data_offset {
            return Err(EngineError::Inference(
                "SafeTensors header exceeds buffer length".into(),
            ));
        }

        let header_json = std::str::from_utf8(&bytes[8..data_offset]).map_err(|e| {
            EngineError::Inference(format!("SafeTensors header not valid UTF-8: {}", e))
        })?;

        let tensors: HashMap<String, SafeTensorEntry> = serde_json::from_str(header_json)
            .map_err(|e| EngineError::Inference(format!("SafeTensors JSON invalid: {}", e)))?;

        Ok(Self {
            tensors,
            data_offset,
        })
    }
}

/// A lightweight, real linear projection model engine for fast embeddings and head inference.
///
/// Multiplies an input token embedding against an output projection matrix `W` of shape `[vocab_size, hidden_dim]`.
/// Supports direct loading of weights for speculative drafting.
#[derive(Debug, Clone)]
pub struct LinearProjectionModel {
    pub name: String,
    pub vocab_size: usize,
    pub hidden_dim: usize,
    /// Weights matrix of shape `[vocab_size, hidden_dim]` in row-major order.
    pub weights: Vec<f32>,
    /// Embedding lookup table `[vocab_size, hidden_dim]`.
    pub embeddings: Vec<f32>,
}

impl LinearProjectionModel {
    /// Creates a new projection model with initial weights.
    pub fn new(name: impl Into<String>, vocab_size: usize, hidden_dim: usize) -> Self {
        let size = vocab_size * hidden_dim;
        let mut weights = Vec::with_capacity(size);
        let mut embeddings = Vec::with_capacity(size);

        // Simple normalized initialization
        let inv = 1.0 / (hidden_dim as f32).sqrt();
        for i in 0..size {
            let val = ((i % 17) as f32 - 8.0) * inv * 0.1;
            weights.push(val);
            embeddings.push(val);
        }

        Self {
            name: name.into(),
            vocab_size,
            hidden_dim,
            weights,
            embeddings,
        }
    }

    /// Computes logits from hidden states via linear projection: `logits = hidden * W^T`.
    pub fn project_hidden(&self, hidden: &[f32]) -> Logits {
        assert_eq!(
            hidden.len(),
            self.hidden_dim,
            "Hidden dimension mismatch in projection"
        );
        let mut logits = Vec::with_capacity(self.vocab_size);

        for row in 0..self.vocab_size {
            let offset = row * self.hidden_dim;
            let mut dot = 0.0f32;
            for (col, &h) in hidden.iter().enumerate().take(self.hidden_dim) {
                dot += h * self.weights[offset + col];
            }
            logits.push(dot);
        }

        Logits::new(logits)
    }

    /// Computes logits directly for an input token using its embedded representation.
    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 LinearProjectionModel {
    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::*;
    use std::io::Cursor;

    #[test]
    fn test_linear_projection_model_forward() {
        let model = LinearProjectionModel::new("test-linear", 16, 8);
        assert_eq!(model.vocab_size, 16);
        assert_eq!(model.hidden_dim, 8);

        let logits = model.project_token(3);
        assert_eq!(logits.vocab_size(), 16);
    }

    #[test]
    fn test_safetensors_metadata_parser() {
        let json_header =
            r#"{"lm_head.weight":{"dtype":"F32","shape":[16,8],"data_offsets":[0,512]}}"#;
        let mut buffer = Vec::new();
        let header_len = json_header.len() as u64;
        buffer.extend_from_slice(&header_len.to_le_bytes());
        buffer.extend_from_slice(json_header.as_bytes());
        buffer.resize(buffer.len() + 512, 0u8);

        let parsed = SafeTensorsMetadata::parse(&buffer).expect("Must parse SafeTensors");
        assert_eq!(parsed.tensors.len(), 1);
        let entry = &parsed.tensors["lm_head.weight"];
        assert_eq!(entry.dtype, "F32");
        assert_eq!(entry.shape, vec![16, 8]);
        assert_eq!(entry.data_offsets, [0, 512]);
    }

    #[test]
    fn test_gguf_metadata_parser_synthetic() {
        let mut buf = Vec::new();
        // Magic
        buf.extend_from_slice(b"GGUF");
        // Version 3
        buf.extend_from_slice(&3u32.to_le_bytes());
        // Tensor count = 1
        buf.extend_from_slice(&1u64.to_le_bytes());
        // Metadata KV count = 2
        buf.extend_from_slice(&2u64.to_le_bytes());

        // KV 1: "general.architecture" -> String "llama"
        let k1 = "general.architecture";
        buf.extend_from_slice(&(k1.len() as u64).to_le_bytes());
        buf.extend_from_slice(k1.as_bytes());
        buf.extend_from_slice(&8u32.to_le_bytes()); // Type 8 = String
        let v1 = "llama";
        buf.extend_from_slice(&(v1.len() as u64).to_le_bytes());
        buf.extend_from_slice(v1.as_bytes());

        // KV 2: "llama.context_length" -> Uint32 4096
        let k2 = "llama.context_length";
        buf.extend_from_slice(&(k2.len() as u64).to_le_bytes());
        buf.extend_from_slice(k2.as_bytes());
        buf.extend_from_slice(&4u32.to_le_bytes()); // Type 4 = Uint32
        buf.extend_from_slice(&4096u32.to_le_bytes());

        // Tensor 1: "token_embd.weight", n_dims=2, dims=[4096, 32000], type=0, offset=0
        let t1 = "token_embd.weight";
        buf.extend_from_slice(&(t1.len() as u64).to_le_bytes());
        buf.extend_from_slice(t1.as_bytes());
        buf.extend_from_slice(&2u32.to_le_bytes()); // 2 dims
        buf.extend_from_slice(&4096u64.to_le_bytes());
        buf.extend_from_slice(&32000u64.to_le_bytes());
        buf.extend_from_slice(&0u32.to_le_bytes()); // F32
        buf.extend_from_slice(&0u64.to_le_bytes()); // offset

        let mut cursor = Cursor::new(buf);
        let meta = GgufMetadata::parse(&mut cursor).expect("GGUF parse failed");

        assert_eq!(meta.version, 3);
        assert_eq!(meta.tensor_count, 1);
        assert_eq!(meta.get_string("general.architecture"), Some("llama"));
        assert_eq!(meta.get_u64("llama.context_length"), Some(4096));
        assert_eq!(meta.tensors.len(), 1);
        assert_eq!(meta.tensors[0].name, "token_embd.weight");
        assert_eq!(meta.tensors[0].dimensions, vec![4096, 32000]);
    }
}