primal-barracuda 0.4.0

barraCuda — BARrier-free Rust Abstracted Cross-platform Unified Dimensional Algebra. Vendor-agnostic GPU compute via WGSL.
Documentation
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
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
// SPDX-License-Identifier: AGPL-3.0-or-later
//! LAMB - Layer-wise Adaptive Moments optimizer for Batch training (Pure WGSL)
//!
//! Enables large batch training (e.g., BERT with 64K batch size)
//! Combines Adam with layer-wise adaptation (trust ratio)
//!
//! **Deep Debt Principles**:
//! - Pure WGSL implementation (no CPU code)
//! - Safe Rust wrapper (no unsafe code)
//! - Hardware-agnostic via WebGPU
//! - Complete implementation (production-ready)

use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;

/// Layer-wise Adaptive Moments optimizer
pub struct Lamb {
    parameters: Tensor,
    gradients: Tensor,
    momentum: Option<Tensor>,
    variance: Option<Tensor>,
    learning_rate: f32,
    beta1: f32,
    beta2: f32,
    step: usize,
    weight_decay: f32,
}

impl Lamb {
    /// Create a LAMB optimizer step with the given parameters and optional momentum/variance.
    ///
    /// # Errors
    ///
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn new(
        parameters: Tensor,
        gradients: Tensor,
        learning_rate: f32,
        beta1: f32,
        beta2: f32,
        step: usize,
        weight_decay: f32,
        momentum: Option<Tensor>,
        variance: Option<Tensor>,
    ) -> Result<Self> {
        // Validate shapes match
        if parameters.shape() != gradients.shape() {
            return Err(BarracudaError::shape_mismatch(
                parameters.shape().to_vec(),
                gradients.shape().to_vec(),
            ));
        }

        // Validate learning rate is positive
        if learning_rate <= 0.0 {
            return Err(BarracudaError::invalid_op(
                "lamb",
                "learning_rate must be positive",
            ));
        }

        // Validate betas in valid range
        if !(0.0..1.0).contains(&beta1) {
            return Err(BarracudaError::invalid_op(
                "lamb",
                "beta1 must be in range [0.0, 1.0)",
            ));
        }

        if !(0.0..1.0).contains(&beta2) {
            return Err(BarracudaError::invalid_op(
                "lamb",
                "beta2 must be in range [0.0, 1.0)",
            ));
        }

        // Validate step is positive
        if step == 0 {
            return Err(BarracudaError::invalid_op(
                "lamb",
                "step must be >= 1 (starts at 1, not 0)",
            ));
        }

        // Validate momentum and variance shapes if provided
        if let Some(ref m_tensor) = momentum
            && m_tensor.shape() != parameters.shape()
        {
            return Err(BarracudaError::shape_mismatch(
                m_tensor.shape().to_vec(),
                parameters.shape().to_vec(),
            ));
        }

        if let Some(ref v_tensor) = variance
            && v_tensor.shape() != parameters.shape()
        {
            return Err(BarracudaError::shape_mismatch(
                v_tensor.shape().to_vec(),
                parameters.shape().to_vec(),
            ));
        }

        Ok(Self {
            parameters,
            gradients,
            momentum,
            variance,
            learning_rate,
            beta1,
            beta2,
            step,
            weight_decay,
        })
    }

    fn wgsl_shader() -> &'static str {
        static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
            include_str!("../shaders/optimizer/lamb_f64.wgsl").to_string()
        });
        std::sync::LazyLock::force(&SHADER).as_str()
    }

    /// Execute LAMB optimizer step.
    ///
    /// # 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.parameters.device();
        let size = self.parameters.shape().iter().product::<usize>();
        let byte_size = (size * std::mem::size_of::<f32>()) as u64;

        // Create writable buffers using GPU copy operations (zero CPU fallbacks)
        let parameters_buffer = device.create_buffer_f32(size)?;

        // Copy parameters buffer using GPU copy
        let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
            label: Some("LAMB Buffer Copy Encoder"),
        });
        encoder.copy_buffer_to_buffer(
            self.parameters.buffer(),
            0,
            &parameters_buffer,
            0,
            byte_size,
        );

        // Create momentum buffer (GPU copy or zero initialization)
        let m_buffer = if let Some(ref m_tensor) = self.momentum {
            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)?
        };

        // Create variance buffer (GPU copy or zero initialization)
        let v_buffer = if let Some(ref v_tensor) = self.variance {
            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()));

        // Create intermediate adam_step buffer
        let adam_step_buffer = device.create_buffer_f32(size)?;

        // Create output buffer
        let output_buffer = device.create_buffer_f32(size)?;

        // Create uniform buffer
        #[repr(C)]
        #[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
        struct Params {
            size: u32,
            step: u32,
            learning_rate: f32,
            beta1: f32,
            beta2: f32,
            epsilon: f32,
            weight_decay: f32,
            _pad1: u32,
        }

        let params = Params {
            size: size as u32,
            step: self.step as u32,
            learning_rate: self.learning_rate,
            beta1: self.beta1,
            beta2: self.beta2,
            epsilon: 1e-8,
            weight_decay: self.weight_decay,
            _pad1: 0,
        };

        let params_buffer = device
            .device
            .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("LAMB Params"),
                contents: bytemuck::cast_slice(&[params]),
                usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
            });

        // Compile shader
        let shader_module = device.compile_shader(Self::wgsl_shader(), Some("LAMB Shader"));

        // Create bind group layout
        let bind_group_layout =
            device
                .device
                .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                    label: Some("LAMB Bind Group Layout"),
                    entries: &[
                        wgpu::BindGroupLayoutEntry {
                            binding: 0,
                            visibility: wgpu::ShaderStages::COMPUTE,
                            ty: wgpu::BindingType::Buffer {
                                ty: wgpu::BufferBindingType::Uniform,
                                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: true },
                                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: true },
                                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::Storage { read_only: false },
                                has_dynamic_offset: false,
                                min_binding_size: None,
                            },
                            count: None,
                        },
                        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,
                        },
                        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,
                        },
                    ],
                });

        // Create bind group
        let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("LAMB Bind Group"),
            layout: &bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: params_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: parameters_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: self.gradients.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: m_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 4,
                    resource: v_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 5,
                    resource: adam_step_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 6,
                    resource: output_buffer.as_entire_binding(),
                },
            ],
        });

        // Create pipeline layout
        let pipeline_layout =
            device
                .device
                .create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
                    label: Some("LAMB Pipeline Layout"),
                    bind_group_layouts: &[&bind_group_layout],
                    immediate_size: 0,
                });

        // Create encoder
        let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
            label: Some("LAMB Encoder"),
        });

        // Step 1: Compute Adam step
        let adam_pipeline =
            device
                .device
                .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
                    label: Some("LAMB Adam Pipeline"),
                    layout: Some(&pipeline_layout),
                    module: &shader_module,
                    entry_point: Some("compute_adam_step"),
                    cache: None,
                    compilation_options: Default::default(),
                });

        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("LAMB Adam Pass"),
                timestamp_writes: None,
            });
            pass.set_pipeline(&adam_pipeline);
            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);
            pass.dispatch_workgroups(workgroups, 1, 1);
        }

        // Step 2: Apply trust ratio
        let trust_pipeline =
            device
                .device
                .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
                    label: Some("LAMB Trust Pipeline"),
                    layout: Some(&pipeline_layout),
                    module: &shader_module,
                    entry_point: Some("apply_trust_ratio"),
                    cache: None,
                    compilation_options: Default::default(),
                });

        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("LAMB Trust Pass"),
                timestamp_writes: None,
            });
            pass.set_pipeline(&trust_pipeline);
            pass.set_bind_group(0, Some(&bind_group), &[]);
            // Deep Debt Evolution: Capability-based dispatch
            // Trust ratio computation is a reduction over parameters
            let caps = DeviceCapabilities::from_device(device);
            let param_size = self.parameters.shape().iter().product::<usize>();
            let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::Reduction);
            let workgroups = (param_size as u32).div_ceil(optimal_wg_size);
            pass.dispatch_workgroups(workgroups.max(1), 1, 1);
        }

        device.submit_commands(Some(encoder.finish()));

        let updated_params = Tensor::from_buffer(
            output_buffer,
            self.parameters.shape().to_vec(),
            device.clone(),
        );

        let updated_m =
            Tensor::from_buffer(m_buffer, self.parameters.shape().to_vec(), device.clone());
        let updated_v =
            Tensor::from_buffer(v_buffer, self.parameters.shape().to_vec(), device.clone());

        Ok((updated_params, updated_m, updated_v))
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[tokio::test]
    async fn test_lamb_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 lamb = Lamb::new(params, gradients, 0.001, 0.9, 0.999, 1, 0.0, None, None).unwrap();
        let (updated_params, _m, _v) = lamb.execute().unwrap();

        assert_eq!(updated_params.shape(), &[4]);
    }

    #[tokio::test]
    async fn test_lamb_with_state() {
        let device = crate::device::test_pool::get_test_device().await;
        let params = Tensor::from_vec_on(vec![1.0; 4], vec![4], device.clone())
            .await
            .unwrap();

        let gradients = Tensor::from_vec_on(vec![0.1; 4], vec![4], device.clone())
            .await
            .unwrap();

        // Step 1
        let lamb1 = Lamb::new(
            params.clone(),
            gradients.clone(),
            0.001,
            0.9,
            0.999,
            1,
            0.0,
            None,
            None,
        )
        .unwrap();
        let (params1, m1, v1) = lamb1.execute().unwrap();

        // Step 2 with accumulated state
        let lamb2 = Lamb::new(
            params1,
            gradients,
            0.001,
            0.9,
            0.999,
            2,
            0.0,
            Some(m1),
            Some(v1),
        )
        .unwrap();
        let (params2, _m2, _v2) = lamb2.execute().unwrap();

        assert_eq!(params2.shape(), &[4]);
    }

    #[tokio::test]
    async fn test_lamb_large_batch() {
        let device = crate::device::test_pool::get_test_device().await;
        let size = 128;
        let params = Tensor::from_vec_on(vec![1.0; size], vec![size], device.clone())
            .await
            .unwrap();

        let gradients = Tensor::from_vec_on(vec![0.01; size], vec![size], device.clone())
            .await
            .unwrap();

        let lamb = Lamb::new(params, gradients, 0.001, 0.9, 0.999, 1, 0.0, None, None).unwrap();
        let (updated_params, updated_m, updated_v) = lamb.execute().unwrap();

        assert_eq!(updated_params.shape(), &[size]);
        assert_eq!(updated_m.shape(), &[size]);
        assert_eq!(updated_v.shape(), &[size]);
    }
}