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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! NAdam Optimizer - GPU-accelerated Nesterov Adam
//!
//! **Deep Debt Principles**:
//! - ✅ Reused existing WGSL shader (smart evolution!)
//! - ✅ Safe Rust wrapper (no unsafe code)
//! - ✅ Hardware-agnostic via WebGPU
//! - ✅ Complete implementation (production-ready training)
//! - ✅ Capability-based dispatch (vendor-optimized workgroups)
//!
//! ## Algorithm
//!
//! NAdam combines Nesterov momentum with Adam's adaptive learning:
//! ```text
//! m_t = β₁*m_{t-1} + (1-β₁)*g_t
//! v_t = β₂*v_{t-1} + (1-β₂)*g_t²
//! m̂_t = (β₁*m_t + (1-β₁)*g_t) / (1-β₁^t) // Nesterov lookahead
//! v̂_t = v_t / (1-β₂^t)
//! θ_t = θ_{t-1} - lr * m̂_t / (√v̂_t + ε)
//! ```
//!
//! **Advantage**: Often faster convergence than vanilla Adam
//!
//! **Reference**: Dozat, 2016 - "Incorporating Nesterov Momentum into Adam"
use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
/// NAdam parameters for WGSL shader
#[repr(C)]
#[derive(Copy, Clone, Debug, bytemuck::Pod, bytemuck::Zeroable)]
struct NAdamParams {
num_params: u32,
learning_rate: f32,
beta1: f32,
beta2: f32,
epsilon: f32,
step: u32,
_padding: [u32; 2],
}
/// NAdam optimizer operation
pub struct NAdam {
params: Tensor,
gradients: Tensor,
m: Option<Tensor>,
v: Option<Tensor>,
learning_rate: f32,
beta1: f32,
beta2: f32,
step: usize,
}
impl NAdam {
/// Create new NAdam optimizer step
pub fn new(
params: Tensor,
gradients: Tensor,
learning_rate: f32,
beta1: f32,
beta2: 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(),
));
}
Ok(Self {
params,
gradients,
m,
v,
learning_rate,
beta1,
beta2,
step,
})
}
/// WGSL shader source
fn shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!(
"../shaders/optimizer/nadam_f64.wgsl"
).to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
}
/// Execute NAdam step (returns updated params, m, v)
///
/// # 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, Tensor, Tensor)> {
let device = self.params.device();
let size = self.params.len();
// Create parameters
let params = NAdamParams {
num_params: size as u32,
learning_rate: self.learning_rate,
beta1: self.beta1,
beta2: self.beta2,
epsilon: 1e-8,
step: self.step as u32,
_padding: [0, 0],
};
let params_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("NAdam Params"),
size: std::mem::size_of::<NAdamParams>() 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));
// Create or get m/v buffers
let zeros = vec![0.0f32; size];
use wgpu::util::DeviceExt;
let m_buffer = if let Some(ref m_tensor) = self.m {
m_tensor.buffer().clone()
} else {
device.device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("nadam_m_init"),
contents: bytemuck::cast_slice(&zeros),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
})
};
let v_buffer = if let Some(ref v_tensor) = self.v {
v_tensor.buffer().clone()
} else {
device.device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("nadam_v_init"),
contents: bytemuck::cast_slice(&zeros),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
})
};
// Output buffers
let params_out_buffer = device.create_buffer_f32(size)?;
let m_out_buffer = device.create_buffer_f32(size)?;
let v_out_buffer = device.create_buffer_f32(size)?;
// Compile shader
let shader = device.compile_shader(Self::shader(), Some("NAdam"));
// Create bind group layout
let bgl = device.device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("NAdam BGL"),
entries: &[
// Gradients (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,
},
// Params (input)
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,
},
// M (input/output)
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,
},
// V (input/output)
wgpu::BindGroupLayoutEntry {
binding: 3,
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,
},
// Params out
wgpu::BindGroupLayoutEntry {
binding: 4,
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,
},
// M out
wgpu::BindGroupLayoutEntry {
binding: 5,
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,
},
// V out
wgpu::BindGroupLayoutEntry {
binding: 6,
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 uniform
wgpu::BindGroupLayoutEntry {
binding: 7,
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("NAdam BG"),
layout: &bgl,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.gradients.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.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: params_out_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 5,
resource: m_out_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 6,
resource: v_out_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 7,
resource: params_buffer.as_entire_binding(),
},
],
});
// Create pipeline
let pipeline_layout = device.device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("NAdam Pipeline Layout"),
bind_group_layouts: &[&bgl],
immediate_size: 0,
});
let pipeline = device.device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("NAdam 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("NAdam Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("NAdam Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
let caps = DeviceCapabilities::from_device(&device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::ElementWise);
let workgroups = (size as u32 + optimal_wg_size - 1) / optimal_wg_size;
pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
// Return updated tensors
let params_shape = self.params.shape().to_vec();
Ok((
Tensor::from_buffer(params_out_buffer, params_shape.clone(), device.clone()),
Tensor::from_buffer(m_out_buffer, params_shape.clone(), device.clone()),
Tensor::from_buffer(v_out_buffer, params_shape, device.clone()),
))
}
}
// ═══════════════════════════════════════════════════════════════
// TENSOR API INTEGRATION
// ═══════════════════════════════════════════════════════════════
impl Tensor {
/// NAdam optimizer step (Nesterov-accelerated Adam)
///
/// **Deep Debt**: Faster convergence than vanilla Adam
///
/// # Arguments
/// - `gradients`: Gradients tensor
/// - `learning_rate`: Learning rate (typically 0.001)
/// - `beta1`: First moment decay (typically 0.9)
/// - `beta2`: Second moment decay (typically 0.999)
/// - `step`: Current training step
/// - `m`: First moment state (None for initialization)
/// - `v`: Second moment state (None for initialization)
///
/// # Returns
/// - (updated_params, updated_m, updated_v)
///
/// # Example
/// ```rust,ignore
/// let (params, m, v) = params.nadam_step(
/// &grads, 0.001, 0.9, 0.999, step,
/// Some(&m), Some(&v)
/// )?;
/// ```
pub fn nadam_step(
self,
gradients: &Self,
learning_rate: f32,
beta1: f32,
beta2: f32,
step: usize,
m: Option<&Self>,
v: Option<&Self>,
) -> Result<(Self, Self, Self)> {
NAdam::new(
self,
gradients.clone(),
learning_rate,
beta1,
beta2,
step,
m.cloned(),
v.cloned(),
)?
.execute()
}
}
// ═══════════════════════════════════════════════════════════════
// TESTS
// ═══════════════════════════════════════════════════════════════
#[cfg(test)]
mod tests {
use super::*;
use crate::device::test_pool::get_test_device;
#[tokio::test]
async fn test_nadam_gpu_basic() {
let device = crate::device::test_pool::get_test_device().await;
let params = Tensor::from_vec_on(vec![1.0; 100], vec![100], device.clone())
.await
.unwrap();
let grads = Tensor::from_vec_on(vec![0.01; 100], vec![100], device)
.await
.unwrap();
let (new_params, m, v) = params.nadam_step(&grads, 0.001, 0.9, 0.999, 1, None, None).unwrap();
assert_eq!(new_params.shape(), &[100]);
assert_eq!(m.shape(), &[100]);
assert_eq!(v.shape(), &[100]);
let data = new_params.to_vec().unwrap();
assert!(data.iter().all(|&x| x.is_finite()));
// Parameters should decrease with positive gradients
assert!(data[0] < 1.0);
}
#[tokio::test]
async fn test_nadam_gpu_with_state() {
let device = crate::device::test_pool::get_test_device().await;
let params = Tensor::from_vec_on(vec![1.0; 10], vec![10], device.clone())
.await
.unwrap();
let grads = Tensor::from_vec_on(vec![0.1; 10], vec![10], device.clone())
.await
.unwrap();
// First step
let (params1, m1, v1) = params.nadam_step(&grads, 0.01, 0.9, 0.999, 1, None, None).unwrap();
// Second step with state
let (params2, _m2, _v2) = params1.nadam_step(&grads, 0.01, 0.9, 0.999, 2, Some(&m1), Some(&v1)).unwrap();
let data = params2.to_vec().unwrap();
assert!(data.iter().all(|&x| x.is_finite()));
}
#[tokio::test]
async fn test_nadam_gpu_convergence() {
let device = crate::device::test_pool::get_test_device().await;
let mut params = Tensor::from_vec_on(vec![10.0; 50], vec![50], device.clone())
.await
.unwrap();
let grads = Tensor::from_vec_on(vec![0.1; 50], vec![50], device)
.await
.unwrap();
let mut m = None;
let mut v = None;
// Multiple steps should decrease parameters
for step in 1..=5 {
let (new_params, new_m, new_v) = params.nadam_step(&grads, 0.01, 0.9, 0.999, step, m.as_ref(), v.as_ref()).unwrap();
params = new_params;
m = Some(new_m);
v = Some(new_v);
}
let final_data = params.to_vec().unwrap();
assert!(final_data[0] < 10.0); // Should have decreased
}
}