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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Adafactor Optimizer - GPU-accelerated Memory-Efficient Adaptive Learning Rates
//!
//! **Deep Debt Principles**:
//! - ✅ Pure WGSL implementation
//! - ✅ Safe Rust wrapper (no unsafe code)
//! - ✅ Hardware-agnostic via WebGPU
//! - ✅ Complete implementation (production-ready)
//! - ✅ Modern idiomatic Rust (no traits, direct impl)
//!
//! Memory-efficient adaptive learning rate optimizer
//! Reduces memory by factorizing second moment matrix
//!
//! Reference: "Adafactor: Adaptive Learning Rates with Sublinear Memory Cost" by Shazeer & Stern (2018)
use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
#[repr(C)]
#[derive(Copy, Clone, Debug, bytemuck::Pod, bytemuck::Zeroable)]
struct AdafactorParams {
size: u32,
lr: f32,
beta1: f32,
beta2: f32,
epsilon1: f32,
epsilon2: f32,
clip_threshold: f32,
decay_rate: f32,
step: u32,
}
/// Adafactor optimizer — memory-efficient adaptive learning rates.
pub struct Adafactor {
gradients: Tensor,
params: Tensor,
m: Option<Tensor>,
v: Option<Tensor>,
learning_rate: f32,
beta1: f32,
beta2: f32,
epsilon1: f32,
epsilon2: f32,
clip_threshold: f32,
decay_rate: f32,
step: usize,
}
impl Adafactor {
/// Create Adafactor optimizer with factorized second moment.
/// # Errors
/// Returns [`Err`] if params and gradients shapes differ, `learning_rate` is not positive,
/// beta1 or beta2 are not in [0.0, 1.0), step is 0, or m/v (if provided) shapes differ from params.
pub fn new(
params: Tensor,
gradients: Tensor,
learning_rate: f32,
beta1: f32,
beta2: f32,
epsilon1: f32,
epsilon2: f32,
clip_threshold: f32,
decay_rate: f32,
step: usize,
m: Option<Tensor>,
v: Option<Tensor>,
) -> Result<Self> {
// Validate shapes match
if params.shape() != gradients.shape() {
return Err(BarracudaError::shape_mismatch(
params.shape().to_vec(),
gradients.shape().to_vec(),
));
}
// Validate learning rate is positive
if learning_rate <= 0.0 {
return Err(BarracudaError::invalid_op(
"adafactor",
"learning_rate must be positive",
));
}
// Validate betas in valid range
if !(0.0..1.0).contains(&beta1) {
return Err(BarracudaError::invalid_op(
"adafactor",
"beta1 must be in range [0.0, 1.0)",
));
}
if !(0.0..1.0).contains(&beta2) {
return Err(BarracudaError::invalid_op(
"adafactor",
"beta2 must be in range [0.0, 1.0)",
));
}
// Validate step is positive
if step == 0 {
return Err(BarracudaError::invalid_op(
"adafactor",
"step must be >= 1 (starts at 1, not 0)",
));
}
// Validate m and v shapes if provided
if let Some(ref m_tensor) = m
&& m_tensor.shape() != params.shape()
{
return Err(BarracudaError::shape_mismatch(
m_tensor.shape().to_vec(),
params.shape().to_vec(),
));
}
if let Some(ref v_tensor) = v
&& v_tensor.shape() != params.shape()
{
return Err(BarracudaError::shape_mismatch(
v_tensor.shape().to_vec(),
params.shape().to_vec(),
));
}
Ok(Self {
gradients,
params,
m,
v,
learning_rate,
beta1,
beta2,
epsilon1,
epsilon2,
clip_threshold,
decay_rate,
step,
})
}
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/optimizer/adafactor_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
}
/// Execute Adafactor step. Returns (`updated_params`, m, v).
/// # Errors
/// Returns [`Err`] if buffer allocation fails, GPU dispatch fails, or the device is lost.
pub fn execute(self) -> Result<(Tensor, Tensor, Tensor)> {
let device = self.params.device();
let size = self.params.shape().iter().product::<usize>();
let adafactor_params = AdafactorParams {
size: size as u32,
lr: self.learning_rate,
beta1: self.beta1,
beta2: self.beta2,
epsilon1: self.epsilon1,
epsilon2: self.epsilon2,
clip_threshold: self.clip_threshold,
decay_rate: self.decay_rate,
step: self.step as u32,
};
// Create writable buffers using GPU copy operations (zero CPU fallbacks)
let byte_size = (size * std::mem::size_of::<f32>()) as u64;
// Copy params to writable buffer using GPU copy
let params_buffer = device.create_buffer_f32(size)?;
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("Adafactor Buffer Copy Encoder"),
});
encoder.copy_buffer_to_buffer(self.params.buffer(), 0, ¶ms_buffer, 0, byte_size);
// Copy or create m buffer (GPU copy or zero initialization)
let m_buffer = if let Some(ref m_tensor) = self.m {
let m_buf = device.create_buffer_f32(size)?;
encoder.copy_buffer_to_buffer(m_tensor.buffer(), 0, &m_buf, 0, byte_size);
m_buf
} else {
device.create_buffer_f32(size)?
};
// Copy or create v buffer (GPU copy or zero initialization)
let v_buffer = if let Some(ref v_tensor) = self.v {
let v_buf = device.create_buffer_f32(size)?;
encoder.copy_buffer_to_buffer(v_tensor.buffer(), 0, &v_buf, 0, byte_size);
v_buf
} else {
device.create_buffer_f32(size)?
};
// Submit buffer copies
device.submit_commands(Some(encoder.finish()));
let adafactor_params_buffer =
device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("adafactor_params"),
contents: bytemuck::cast_slice(&[adafactor_params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("adafactor_shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("adafactor_bind_group_layout"),
entries: &[
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,
},
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,
},
wgpu::BindGroupLayoutEntry {
binding: 2,
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,
},
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,
},
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,
},
],
});
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("adafactor_pipeline_layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("adafactor_pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("adafactor_bind_group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.gradients.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: params_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: m_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: v_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: adafactor_params_buffer.as_entire_binding(),
},
],
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("adafactor_encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("adafactor_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::ElementWise);
let workgroups = (size as u32).div_ceil(optimal_wg_size);
compute_pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
let updated_params =
Tensor::from_buffer(params_buffer, self.params.shape().to_vec(), device.clone());
let updated_m = Tensor::from_buffer(m_buffer, self.params.shape().to_vec(), device.clone());
let updated_v = Tensor::from_buffer(v_buffer, self.params.shape().to_vec(), device.clone());
Ok((updated_params, updated_m, updated_v))
}
}
impl Tensor {
/// Adafactor optimizer step
/// # Arguments
/// - `gradients`: Gradient tensor [same shape as params]
/// - `learning_rate`: Learning rate
/// - `beta1`: Exponential decay for first moment (0 = disable), typically 0.0
/// - `beta2`: Exponential decay for second moment, typically 0.999
/// - `epsilon1`: Regularization for RMS, typically 1e-30
/// - `epsilon2`: Regularization for parameter scale, typically 1e-3
/// - `clip_threshold`: Gradient clipping threshold, typically 1.0
/// - `decay_rate`: Decay rate, typically -0.8
/// - `step`: Current iteration (starts at 1, not 0)
/// - `m`: First moment estimate (None for first step or if beta1=0)
/// - `v`: Second moment estimate (None for first step)
/// # Returns
/// - Tuple: (`updated_params`, `updated_m`, `updated_v`)
/// # Errors
/// Returns [`Err`] if params and gradients shapes differ, `learning_rate` is not positive,
/// beta1 or beta2 are not in [0.0, 1.0), step is 0, m/v shapes differ, buffer allocation fails,
/// GPU dispatch fails, or the device is lost.
pub fn adafactor_step(
self,
gradients: &Self,
learning_rate: f32,
beta1: f32,
beta2: f32,
epsilon1: f32,
epsilon2: f32,
clip_threshold: f32,
decay_rate: f32,
step: usize,
m: Option<&Self>,
v: Option<&Self>,
) -> Result<(Self, Self, Self)> {
Adafactor::new(
self,
gradients.clone(),
learning_rate,
beta1,
beta2,
epsilon1,
epsilon2,
clip_threshold,
decay_rate,
step,
m.cloned(),
v.cloned(),
)?
.execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_adafactor_basic() {
let device = crate::device::test_pool::get_test_device().await;
let params = Tensor::from_vec_on(vec![1.0, 2.0, 3.0, 4.0], vec![4], device.clone())
.await
.unwrap();
let gradients = Tensor::from_vec_on(vec![0.1, 0.2, 0.3, 0.4], vec![4], device.clone())
.await
.unwrap();
let (updated_params, _m, _v) = params
.adafactor_step(
&gradients, 0.001, 0.0, 0.999, 1e-30, 1e-3, 1.0, -0.8, 1, None, None,
)
.unwrap();
let result = updated_params.to_vec().unwrap();
assert_eq!(result.len(), 4);
assert!(result.iter().all(|&x| x.is_finite()));
}
}