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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Rotary Position Embedding (`RoPE`) - GPU-accelerated
//!
//! **Deep Debt Principles**:
//! - ✅ Pure WGSL implementation (single-pass GPU)
//! - ✅ Safe Rust wrapper (no unsafe code)
//! - ✅ Hardware-agnostic via WebGPU
//! - ✅ Complete implementation (production-ready for Llama, GPT-NeoX)
//!
//! ## Algorithm
//!
//! ```text
//! RoPE applies 2D rotations to embedding pairs based on position:
//!
//! freq[i] = 1 / (10000^(2i / head_dim))
//! theta[pos,i] = pos * freq[i]
//! [x1', x2'] = [cos(theta) -sin(theta)] [x1]
//! [sin(theta) cos(theta)] [x2]
//! ```
//!
//! **Key Properties**:
//! - Encodes relative position (not absolute)
//! - No learned parameters
//! - Rotation preserves magnitude
//! - Works for any sequence length
//!
//! **Used By**: Llama, GPT-NeoX, `PaLM`, Falcon
//!
//! **Reference**: `RoFormer` (Su et al., 2021)
//!
//! ## Usage
//!
//! ```rust,ignore
//! use barracuda::tensor::Tensor;
//!
//! let q = Tensor::randn(vec![2, 128, 8, 64]).await?; // [batch, seq, heads, dim]
//! let q_rotated = q.rotary_embedding()?; // Apply RoPE
//! ```
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/rotary_embedding_f64.wgsl");
static SHADER_F32: std::sync::LazyLock<String> =
std::sync::LazyLock::new(|| SHADER_F64.to_string());
/// `RoPE` parameters for WGSL shader
#[repr(C)]
#[derive(Copy, Clone, Debug, bytemuck::Pod, bytemuck::Zeroable)]
struct RopeParams {
batch_size: u32,
seq_len: u32,
num_heads: u32,
head_dim: u32,
half_dim: u32,
_padding: [u32; 3],
}
/// Rotary Position Embedding operation
///
/// **Deep Debt**: Single-pass GPU implementation, no dependencies
pub struct RotaryEmbedding {
input: Tensor,
}
impl RotaryEmbedding {
/// Create new `RoPE` operation
/// **Shape**: [batch, `seq_len`, `num_heads`, `head_dim`]
/// # Errors
/// Returns [`Err`] if input is not 4D or `head_dim` is not even.
pub fn new(input: Tensor) -> Result<Self> {
// Validate shape: must be 4D
if input.shape().len() != 4 {
return Err(BarracudaError::shape_mismatch(
input.shape().to_vec(),
vec![0, 0, 0, 0],
));
}
// Validate head_dim is even (required for pairwise rotation)
let head_dim = input.shape()[3];
if !head_dim.is_multiple_of(2) {
return Err(BarracudaError::invalid_op(
"RotaryEmbedding",
format!("head_dim ({head_dim}) must be even for pairwise rotation"),
));
}
Ok(Self { input })
}
/// WGSL shader source
fn shader() -> &'static str {
&SHADER_F32
}
/// Execute `RoPE` (single GPU pass)
/// **Deep Debt**: Efficient single-pass implementation
/// # Errors
/// Returns [`Err`] if buffer allocation, GPU dispatch, or device submission fails (e.g. device lost).
pub fn execute(self) -> Result<Tensor> {
let device = self.input.device();
// Extract dimensions
let shape = self.input.shape();
let batch_size = shape[0];
let seq_len = shape[1];
let num_heads = shape[2];
let head_dim = shape[3];
let half_dim = head_dim / 2;
// Create parameters
let params = RopeParams {
batch_size: batch_size as u32,
seq_len: seq_len as u32,
num_heads: num_heads as u32,
head_dim: head_dim as u32,
half_dim: half_dim as u32,
_padding: [0, 0, 0],
};
let params_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("RoPE Params"),
size: std::mem::size_of::<RopeParams>() as u64,
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
mapped_at_creation: false,
});
device
.queue
.write_buffer(¶ms_buffer, 0, bytemuck::bytes_of(¶ms));
// Output buffer (same size as input)
let output_size = self.input.len();
let output_buffer = device.create_buffer_f32(output_size)?;
// Compile shader
let shader = device.compile_shader(Self::shader(), Some("RoPE"));
// Create bind group layout
let bgl = device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("RoPE BGL"),
entries: &[
// Input
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,
},
// Output
wgpu::BindGroupLayoutEntry {
binding: 1,
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,
},
// Params
wgpu::BindGroupLayoutEntry {
binding: 2,
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("RoPE BG"),
layout: &bgl,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: params_buffer.as_entire_binding(),
},
],
});
// Create pipeline
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("RoPE Pipeline Layout"),
bind_group_layouts: &[&bgl],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("RoPE Pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
// Execute
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("RoPE Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("RoPE Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
// Deep Debt Evolution: Capability-based dispatch
let total = (batch_size * seq_len * num_heads * half_dim) as u32;
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::ElementWise);
let workgroups = total.div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
// Return output tensor (same shape as input)
Ok(Tensor::from_buffer(
output_buffer,
vec![batch_size, seq_len, num_heads, head_dim],
device.clone(),
))
}
}
// ═══════════════════════════════════════════════════════════════
// TENSOR API INTEGRATION
// ═══════════════════════════════════════════════════════════════
impl Tensor {
/// Apply Rotary Position Embedding (`RoPE`)
/// **Deep Debt**: Essential for Llama, GPT-NeoX, `PaLM`
/// # Arguments
/// - Input: [batch, `seq_len`, `num_heads`, `head_dim`]
/// # Returns
/// - Output: [batch, `seq_len`, `num_heads`, `head_dim`] (rotated)
/// # Example
/// ```rust,ignore
/// let q = Tensor::randn(vec![2, 128, 8, 64]).await?;
/// let q_rope = q.rotary_embedding()?; // Apply RoPE for Llama
/// ```
/// # Errors
/// Returns [`Err`] if input is not 4D, `head_dim` is not even, or buffer allocation/GPU dispatch fails (e.g. device lost).
pub fn rotary_embedding(self) -> Result<Self> {
RotaryEmbedding::new(self)?.execute()
}
}
// ═══════════════════════════════════════════════════════════════
// TESTS
// ═══════════════════════════════════════════════════════════════
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_rope_basic() {
let device = crate::device::test_pool::get_test_device().await;
let batch = 1;
let seq = 4;
let heads = 2;
let dim = 8;
let input = Tensor::from_vec_on(
vec![1.0; batch * seq * heads * dim],
vec![batch, seq, heads, dim],
device,
)
.await
.unwrap();
let output = input.rotary_embedding().unwrap();
assert_eq!(output.shape(), &[batch, seq, heads, dim]);
let data = output.to_vec().unwrap();
assert!(data.iter().all(|&x| x.is_finite()));
}
#[tokio::test]
async fn test_rope_single_position() {
let device = crate::device::test_pool::get_test_device().await;
let batch = 1;
let seq = 1;
let heads = 2;
let dim = 8;
let input = Tensor::from_vec_on(
vec![1.0; batch * seq * heads * dim],
vec![batch, seq, heads, dim],
device,
)
.await
.unwrap();
let output = input.rotary_embedding().unwrap();
assert_eq!(output.shape(), &[batch, seq, heads, dim]);
}
#[tokio::test]
async fn test_rope_llama_dims() {
let device = crate::device::test_pool::get_test_device().await;
// Llama-style dimensions
let batch = 2;
let seq = 128;
let heads = 8;
let dim = 64;
let input = Tensor::from_vec_on(
vec![0.5; batch * seq * heads * dim],
vec![batch, seq, heads, dim],
device,
)
.await
.unwrap();
let output = input.rotary_embedding().unwrap();
assert_eq!(output.shape(), &[batch, seq, heads, dim]);
let data = output.to_vec().unwrap();
assert!(data.iter().all(|&x| x.is_finite()));
}
#[tokio::test]
async fn test_rope_magnitude_preservation() {
let device = crate::device::test_pool::get_test_device().await;
let batch = 1;
let seq = 4;
let heads = 1;
let dim = 4;
let input = Tensor::from_vec_on(
vec![1.0; batch * seq * heads * dim],
vec![batch, seq, heads, dim],
device,
)
.await
.unwrap();
let output = input.rotary_embedding().unwrap();
// Rotation should preserve magnitude (approximately)
let data = output.to_vec().unwrap();
assert!(data.iter().all(|&x| x.abs() <= 2.0));
}
#[tokio::test]
async fn test_rope_shape_validation() {
let device = crate::device::test_pool::get_test_device().await;
// Valid: even head_dim
let input = Tensor::from_vec_on(vec![1.0; 4 * 2 * 8], vec![1, 4, 2, 8], device.clone())
.await
.unwrap();
assert!(input.rotary_embedding().is_ok());
// Invalid: odd head_dim
let input = Tensor::from_vec_on(vec![1.0; 4 * 2 * 7], vec![1, 4, 2, 7], device)
.await
.unwrap();
assert!(input.rotary_embedding().is_err());
}
}