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
impl Attention {
/// Create a new attention layer
///
/// # Arguments
///
/// * `head_dim` - Dimension of each attention head
///
/// # Errors
///
/// Returns error if `head_dim` is zero
pub fn new(head_dim: usize) -> Result<Self> {
if head_dim == 0 {
return Err(RealizarError::InvalidShape {
reason: "head_dim must be > 0".to_string(),
});
}
#[allow(clippy::cast_precision_loss)]
let scale = 1.0 / (head_dim as f32).sqrt();
Ok(Self { head_dim, scale })
}
/// Compute scaled dot-product attention
///
/// # Arguments
///
/// * `query` - Query tensor `[seq_len, head_dim]`
/// * `key` - Key tensor `[seq_len, head_dim]`
/// * `value` - Value tensor `[seq_len, head_dim]`
///
/// # Returns
///
/// Output tensor `[seq_len, head_dim]`
///
/// # Errors
///
/// Returns error if shapes don't match
pub fn forward(
&self,
query: &Tensor<f32>,
key: &Tensor<f32>,
value: &Tensor<f32>,
) -> Result<Tensor<f32>> {
let q_shape = query.shape();
let k_shape = key.shape();
let v_shape = value.shape();
// Validate shapes
if q_shape.is_empty() || k_shape.is_empty() || v_shape.is_empty() {
return Err(RealizarError::InvalidShape {
reason: "Query, key, value tensors must have at least 1 dimension".to_string(),
});
}
let q_last = q_shape[q_shape.len() - 1];
let k_last = k_shape[k_shape.len() - 1];
let v_last = v_shape[v_shape.len() - 1];
if q_last != self.head_dim || k_last != self.head_dim || v_last != self.head_dim {
return Err(RealizarError::InvalidShape {
reason: format!(
"Expected head_dim={}, got Q={}, K={}, V={}",
self.head_dim, q_last, k_last, v_last
),
});
}
// Get sequence lengths
let q_seq_len = if q_shape.len() > 1 { q_shape[0] } else { 1 };
let k_seq_len = if k_shape.len() > 1 { k_shape[0] } else { 1 };
let v_seq_len = if v_shape.len() > 1 { v_shape[0] } else { 1 };
if k_seq_len != v_seq_len {
return Err(RealizarError::InvalidShape {
reason: format!("Key seq_len {k_seq_len} != Value seq_len {v_seq_len}"),
});
}
let q_data = query.data();
let k_data = key.data();
let v_data = value.data();
// Compute attention scores: Q @ K.T
// scores[i][j] = sum(Q[i][k] * K[j][k]) for all k
let mut scores = Vec::with_capacity(q_seq_len * k_seq_len);
for i in 0..q_seq_len {
for j in 0..k_seq_len {
let mut dot = 0.0;
for k in 0..self.head_dim {
dot += q_data[i * self.head_dim + k] * k_data[j * self.head_dim + k];
}
scores.push(dot * self.scale);
}
}
// Apply softmax to each row of scores
let scores_tensor = Tensor::from_vec(vec![q_seq_len, k_seq_len], scores)?;
let attn_weights = softmax(&scores_tensor)?;
let attn_data = attn_weights.data();
// Compute output: attn_weights @ V
// output[i][k] = sum(attn_weights[i][j] * V[j][k]) for all j
let mut output = Vec::with_capacity(q_seq_len * self.head_dim);
for i in 0..q_seq_len {
for k in 0..self.head_dim {
let mut sum = 0.0;
for j in 0..k_seq_len {
sum += attn_data[i * k_seq_len + j] * v_data[j * self.head_dim + k];
}
output.push(sum);
}
}
// Debug assertion for numerical stability
debug_assert!(
output.iter().all(|&x| x.is_finite()),
"Attention layer produced NaN or Inf values - check input scaling"
);
Tensor::from_vec(vec![q_seq_len, self.head_dim], output)
}
/// Get head dimension
#[must_use]
pub fn head_dim(&self) -> usize {
self.head_dim
}
/// Get scale factor
#[must_use]
pub fn scale(&self) -> f32 {
self.scale
}
/// Compute Flash Attention - memory-efficient block-wise attention
///
/// Uses tiling and recomputation to reduce memory usage from O(N²) to O(N).
/// Implements block-wise softmax with running max/sum statistics.
///
/// # Arguments
///
/// * `query` - Query tensor `[seq_len, head_dim]`
/// * `key` - Key tensor `[seq_len, head_dim]`
/// * `value` - Value tensor `[seq_len, head_dim]`
/// * `block_size` - Tile size for block-wise computation (e.g., 64, 128)
///
/// # Returns
///
/// Output tensor `[seq_len, head_dim]` (same as standard attention)
///
/// # Errors
///
/// Returns error if shapes don't match or `block_size` is zero
///
/// # References
///
/// - "`FlashAttention`: Fast and Memory-Efficient Exact Attention" - Dao et al., 2022
/// - "FlashAttention-2: Faster Attention with Better Parallelism" - Dao, 2023
pub fn flash_forward(
&self,
query: &Tensor<f32>,
key: &Tensor<f32>,
value: &Tensor<f32>,
block_size: usize,
) -> Result<Tensor<f32>> {
if block_size == 0 {
return Err(RealizarError::InvalidShape {
reason: "block_size must be > 0".to_string(),
});
}
let q_shape = query.shape();
let k_shape = key.shape();
let v_shape = value.shape();
// Validate shapes (same as standard attention)
if q_shape.is_empty() || k_shape.is_empty() || v_shape.is_empty() {
return Err(RealizarError::InvalidShape {
reason: "Query, key, value tensors must have at least 1 dimension".to_string(),
});
}
let q_last = q_shape[q_shape.len() - 1];
let k_last = k_shape[k_shape.len() - 1];
let v_last = v_shape[v_shape.len() - 1];
if q_last != self.head_dim || k_last != self.head_dim || v_last != self.head_dim {
return Err(RealizarError::InvalidShape {
reason: format!(
"Expected head_dim={}, got Q={}, K={}, V={}",
self.head_dim, q_last, k_last, v_last
),
});
}
// Get sequence lengths
let q_seq_len = if q_shape.len() > 1 { q_shape[0] } else { 1 };
let k_seq_len = if k_shape.len() > 1 { k_shape[0] } else { 1 };
let v_seq_len = if v_shape.len() > 1 { v_shape[0] } else { 1 };
if k_seq_len != v_seq_len {
return Err(RealizarError::InvalidShape {
reason: format!("Key seq_len {k_seq_len} != Value seq_len {v_seq_len}"),
});
}
let q_data = query.data();
let k_data = key.data();
let v_data = value.data();
// Initialize output and statistics
let mut output = vec![0.0; q_seq_len * self.head_dim];
let mut row_max = vec![f32::NEG_INFINITY; q_seq_len]; // Running max for each query row
let mut row_sum = vec![0.0; q_seq_len]; // Running sum for each query row
// Iterate over K/V blocks (outer loop)
let num_kv_blocks = k_seq_len.div_ceil(block_size);
for kv_block_idx in 0..num_kv_blocks {
let kv_start = kv_block_idx * block_size;
let kv_end = (kv_start + block_size).min(k_seq_len);
let kv_block_len = kv_end - kv_start;
// Iterate over Q blocks (inner loop)
let num_q_blocks = q_seq_len.div_ceil(block_size);
for q_block_idx in 0..num_q_blocks {
let q_start = q_block_idx * block_size;
let q_end = (q_start + block_size).min(q_seq_len);
// Compute attention scores for this block: Q_block @ K_block.T
let mut scores = vec![0.0; (q_end - q_start) * kv_block_len];
for (i, q_idx) in (q_start..q_end).enumerate() {
for (j, kv_idx) in (kv_start..kv_end).enumerate() {
let mut dot = 0.0;
for k in 0..self.head_dim {
dot += q_data[q_idx * self.head_dim + k]
* k_data[kv_idx * self.head_dim + k];
}
scores[i * kv_block_len + j] = dot * self.scale;
}
}
// Update running max and apply softmax with new max
for (i, q_idx) in (q_start..q_end).enumerate() {
// Find max in current block
let block_max = (0..kv_block_len)
.map(|j| scores[i * kv_block_len + j])
.fold(f32::NEG_INFINITY, f32::max);
// Update global max
let old_max = row_max[q_idx];
let new_max = old_max.max(block_max);
row_max[q_idx] = new_max;
// Compute exp(scores - new_max) and update running sum
let mut block_sum = 0.0;
for j in 0..kv_block_len {
let exp_val = (scores[i * kv_block_len + j] - new_max).exp();
scores[i * kv_block_len + j] = exp_val;
block_sum += exp_val;
}
// Rescale old output and sum based on new max
let scale_factor = (old_max - new_max).exp();
for k in 0..self.head_dim {
output[q_idx * self.head_dim + k] *= scale_factor;
}
row_sum[q_idx] = row_sum[q_idx] * scale_factor + block_sum;
}
// Accumulate weighted values: output += scores @ V_block
for (i, q_idx) in (q_start..q_end).enumerate() {
for k in 0..self.head_dim {
let mut weighted_sum = 0.0;
for (j, kv_idx) in (kv_start..kv_end).enumerate() {
weighted_sum +=
scores[i * kv_block_len + j] * v_data[kv_idx * self.head_dim + k];
}
output[q_idx * self.head_dim + k] += weighted_sum;
}
}
}
}
// Final normalization by row_sum
for i in 0..q_seq_len {
for k in 0..self.head_dim {
output[i * self.head_dim + k] /= row_sum[i];
}
}
Tensor::from_vec(vec![q_seq_len, self.head_dim], output)
}
/// Flash Attention v2 with SIMD-accelerated dot products
///
/// Optimized implementation using AVX2 SIMD for dot products.
/// Uses parallel outer loop over query blocks for better multi-core utilization.
///
/// # Arguments
///
/// * `query` - Query tensor `[seq_len, head_dim]`
/// * `key` - Key tensor `[seq_len, head_dim]`
/// * `value` - Value tensor `[seq_len, head_dim]`
/// * `block_size` - Tile size for block-wise computation (e.g., 64, 128)
///
/// # Returns
///
/// Output tensor `[seq_len, head_dim]` (same as standard attention)
///
/// # Errors
///
/// Returns error if shapes don't match or `block_size` is zero
///
/// # References
///
/// - "FlashAttention-2: Faster Attention with Better Parallelism" - Dao, 2023
#[allow(clippy::similar_names)]
pub fn flash_forward_v2(
&self,
query: &Tensor<f32>,
key: &Tensor<f32>,
value: &Tensor<f32>,
block_size: usize,
) -> Result<Tensor<f32>> {
if block_size == 0 {
return Err(RealizarError::InvalidShape {
reason: "block_size must be > 0".to_string(),
});
}
let q_shape = query.shape();
let k_shape = key.shape();
let v_shape = value.shape();
// Validate shapes
if q_shape.is_empty() || k_shape.is_empty() || v_shape.is_empty() {
return Err(RealizarError::InvalidShape {
reason: "Query, key, value tensors must have at least 1 dimension".to_string(),
});
}
let q_last = q_shape[q_shape.len() - 1];
let k_last = k_shape[k_shape.len() - 1];
let v_last = v_shape[v_shape.len() - 1];
if q_last != self.head_dim || k_last != self.head_dim || v_last != self.head_dim {
return Err(RealizarError::InvalidShape {
reason: format!(
"Expected head_dim={}, got Q={}, K={}, V={}",
self.head_dim, q_last, k_last, v_last
),
});
}
let q_seq_len = if q_shape.len() > 1 { q_shape[0] } else { 1 };
let k_seq_len = if k_shape.len() > 1 { k_shape[0] } else { 1 };
let v_seq_len = if v_shape.len() > 1 { v_shape[0] } else { 1 };
if k_seq_len != v_seq_len {
return Err(RealizarError::InvalidShape {
reason: format!("Key seq_len {k_seq_len} != Value seq_len {v_seq_len}"),
});
}
let q_data = query.data();
let k_data = key.data();
let v_data = value.data();
let head_dim = self.head_dim;
let scale = self.scale;
// Initialize output and statistics
let mut output = vec![0.0; q_seq_len * head_dim];
let mut row_max = vec![f32::NEG_INFINITY; q_seq_len];
let mut row_sum = vec![0.0; q_seq_len];
// Flash Attention v2: Iterate over K/V blocks in outer loop
// This allows better memory access patterns
let num_kv_blocks = k_seq_len.div_ceil(block_size);
for kv_block_idx in 0..num_kv_blocks {
let kv_start = kv_block_idx * block_size;
let kv_end = (kv_start + block_size).min(k_seq_len);
let kv_block_len = kv_end - kv_start;
// Process all Q rows against this K/V block
for q_idx in 0..q_seq_len {
// SIMD-accelerated dot products for this row
let mut scores = Vec::with_capacity(kv_block_len);
for kv_idx in kv_start..kv_end {
let dot = Self::simd_dot_product(
&q_data[q_idx * head_dim..(q_idx + 1) * head_dim],
&k_data[kv_idx * head_dim..(kv_idx + 1) * head_dim],
);
scores.push(dot * scale);
}
// Find max in current block
let block_max = scores.iter().fold(f32::NEG_INFINITY, |a, &b| a.max(b));
// Update global max
let old_max = row_max[q_idx];
let new_max = old_max.max(block_max);
row_max[q_idx] = new_max;
// Compute exp(scores - new_max) and update running sum
let mut block_sum = 0.0;
for score in &mut scores {
let exp_val = (*score - new_max).exp();
*score = exp_val;
block_sum += exp_val;
}
// Rescale old output and sum based on new max
let scale_factor = (old_max - new_max).exp();
for k in 0..head_dim {
output[q_idx * head_dim + k] *= scale_factor;
}
row_sum[q_idx] = row_sum[q_idx] * scale_factor + block_sum;
// Accumulate weighted values: output += scores @ V_block
for (j, kv_idx) in (kv_start..kv_end).enumerate() {
let weight = scores[j];
for k in 0..head_dim {
output[q_idx * head_dim + k] += weight * v_data[kv_idx * head_dim + k];
}
}
}
}
// Final normalization by row_sum
for i in 0..q_seq_len {
let inv_sum = 1.0 / row_sum[i];
for k in 0..head_dim {
output[i * head_dim + k] *= inv_sum;
}
}
Tensor::from_vec(vec![q_seq_len, self.head_dim], output)
}
}