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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Flash Attention - Memory-efficient attention mechanism
//!
//! **Deep Debt Principles**:
//! - ✅ Pure WGSL implementation (hardware-agnostic)
//! - ✅ Safe Rust wrapper (no unsafe code)
//! - ✅ Memory-efficient O(N) instead of O(N²)
//! - ✅ Complete implementation (production-ready)
//!
//! ## Algorithm
//!
//! Flash Attention reduces memory complexity from O(N²) to O(N) by:
//! 1. Computing attention in blocks (tiling)
//! 2. Fusing operations to minimize memory access
//! 3. Using numerically stable softmax
//!
//! Reference: "`FlashAttention`: Fast and Memory-Efficient Exact Attention"
//! by Dao et al. (2022) - 2-4x faster than standard attention!
//!
//! ## Usage
//!
//! ```rust,ignore
//! use barracuda::tensor::Tensor;
//!
//! let query = Tensor::randn(vec![seq_len, head_dim]).await?;
//! let key = Tensor::randn(vec![seq_len, head_dim]).await?;
//! let value = Tensor::randn(vec![seq_len, head_dim]).await?;
//!
//! let output = query.flash_attention(&key, &value, num_heads)?;
//! ```
use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
/// f64 is the canonical source — math is universal, precision is silicon.
static SHADER_F64: &str = include_str!("../shaders/attention/flash_attention_f64.wgsl");
static SHADER_F32: std::sync::LazyLock<String> =
std::sync::LazyLock::new(|| SHADER_F64.to_string());
#[repr(C)]
#[derive(Copy, Clone, Debug, bytemuck::Pod, bytemuck::Zeroable)]
struct FlashAttentionParams {
seq_len: u32,
head_dim: u32,
num_heads: u32,
scale: f32,
}
/// Flash Attention operation
pub struct FlashAttention {
query: Tensor,
key: Tensor,
value: Tensor,
num_heads: u32,
}
impl FlashAttention {
/// Create Flash Attention operation
/// # Errors
/// Returns [`Err`] if shape validation fails (e.g. head dimension mismatch).
pub fn new(query: Tensor, key: Tensor, value: Tensor, num_heads: u32) -> Result<Self> {
// Verify shapes match
if query.shape() != key.shape() || query.shape() != value.shape() {
return Err(BarracudaError::invalid_op(
"flash_attention",
format!(
"Q, K, V must have same shape, got Q: {:?}, K: {:?}, V: {:?}",
query.shape(),
key.shape(),
value.shape()
),
));
}
// Verify 2D shape [seq_len, head_dim]
if query.shape().len() != 2 {
return Err(BarracudaError::invalid_op(
"flash_attention",
format!(
"Requires 2D tensors [seq_len, head_dim], got shape: {:?}",
query.shape()
),
));
}
Ok(Self {
query,
key,
value,
num_heads,
})
}
/// WGSL shader source
fn wgsl_shader() -> &'static str {
&SHADER_F32
}
/// Execute flash attention
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn execute(self) -> Result<Tensor> {
let device = self.query.device();
let shape = self.query.shape();
let seq_len = shape[0];
let head_dim = shape[1];
let size = seq_len * head_dim;
// Create output buffer
let output_buffer = device.create_buffer_f32(size)?;
// Calculate scale factor (1 / sqrt(head_dim))
let scale = 1.0 / (head_dim as f32).sqrt();
// Create parameters
let params = FlashAttentionParams {
seq_len: seq_len as u32,
head_dim: head_dim as u32,
num_heads: self.num_heads,
scale,
};
// Create uniform buffer for parameters
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Flash Attention Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
// Create bind group layout
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("Flash Attention Bind Group Layout"),
entries: &[
// Query buffer
wgpu::BindGroupLayoutEntry {
binding: 0,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Storage { read_only: true },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
// Key buffer
wgpu::BindGroupLayoutEntry {
binding: 1,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Storage { read_only: true },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
// Value buffer
wgpu::BindGroupLayoutEntry {
binding: 2,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Storage { read_only: true },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
// Output buffer
wgpu::BindGroupLayoutEntry {
binding: 3,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Storage { read_only: false },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
// Parameters uniform
wgpu::BindGroupLayoutEntry {
binding: 4,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
// Create bind group
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("Flash Attention Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.query.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.key.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: self.value.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: params_buffer.as_entire_binding(),
},
],
});
// Compile shader
let shader_module =
device.compile_shader(Self::wgsl_shader(), Some("Flash Attention Shader"));
// Create pipeline
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("Flash Attention Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("Flash Attention Pipeline"),
layout: Some(&pipeline_layout),
module: &shader_module,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
// Execute compute pass
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("Flash Attention Encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Flash Attention Pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&pipeline);
compute_pass.set_bind_group(0, Some(&bind_group), &[]);
// Deep Debt Evolution: Capability-based dispatch
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::MatMul);
let total_size = seq_len * head_dim;
let workgroups = (total_size as u32).div_ceil(optimal_wg_size);
compute_pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
// Return result tensor
Ok(Tensor::from_buffer(
output_buffer,
shape.to_vec(),
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_flash_attention_basic() {
let device = crate::device::test_pool::get_test_device().await;
let seq_len = 4;
let head_dim = 8;
// Create simple Q, K, V
let q_data = vec![1.0; seq_len * head_dim];
let k_data = vec![1.0; seq_len * head_dim];
let v_data = vec![2.0; seq_len * head_dim];
let query = Tensor::from_vec_on(q_data, vec![seq_len, head_dim], device.clone())
.await
.unwrap();
let key = Tensor::from_vec_on(k_data, vec![seq_len, head_dim], device.clone())
.await
.unwrap();
let value = Tensor::from_vec_on(v_data, vec![seq_len, head_dim], device)
.await
.unwrap();
let output = FlashAttention::new(query, key, value, 1)
.unwrap()
.execute()
.unwrap();
let result = output.to_vec().unwrap();
assert_eq!(result.len(), seq_len * head_dim);
// Output should be weighted sum of values (all should be close to 2.0)
for &val in &result {
assert!((val - 2.0).abs() < 1.0, "Expected ~2.0, got {val}");
}
}
#[tokio::test]
async fn test_flash_attention_shape_validation() {
let device = crate::device::test_pool::get_test_device().await;
let q = Tensor::from_vec_on(vec![1.0; 16], vec![4, 4], device.clone())
.await
.unwrap();
let k = Tensor::from_vec_on(vec![1.0; 12], vec![3, 4], device.clone())
.await
.unwrap();
let v = Tensor::from_vec_on(vec![1.0; 16], vec![4, 4], device)
.await
.unwrap();
// Should fail: mismatched shapes
let result = FlashAttention::new(q, k, v, 1);
assert!(result.is_err());
}
#[tokio::test]
async fn test_flash_attention_small_sequence() {
let device = crate::device::test_pool::get_test_device().await;
let seq_len = 2;
let head_dim = 4;
let query = Tensor::from_vec_on(
vec![1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0],
vec![seq_len, head_dim],
device.clone(),
)
.await
.unwrap();
let key = Tensor::from_vec_on(
vec![1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0],
vec![seq_len, head_dim],
device.clone(),
)
.await
.unwrap();
let value = Tensor::from_vec_on(
vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0],
vec![seq_len, head_dim],
device,
)
.await
.unwrap();
let output = FlashAttention::new(query, key, value, 1)
.unwrap()
.execute()
.unwrap();
let result = output.to_vec().unwrap();
assert_eq!(result.len(), seq_len * head_dim);
// Each position should attend to all positions (uniform attention)
assert!(result.iter().all(|&x| x.is_finite()));
}
}