entrenar 0.7.10

Training & Optimization library with autograd, LoRA, quantization, and model merging
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

use super::*;
/// FALSIFY-WGSL-GEMM-001: Tiled GEMM matches naive matmul
#[test]
fn test_wgpu_matmul_forward() {
    let trainer = match WgpuTrainer::new() {
        Ok(t) => t,
        Err(_) => return, // Skip if no GPU
    };

    // Small matrix: A[4,8] @ B[8,6] = C[4,6]
    let m = 4u32;
    let k = 8u32;
    let n = 6u32;

    let a_data: Vec<f32> = (0..m * k).map(|i| (i as f32) * 0.1).collect();
    let b_data: Vec<f32> = (0..k * n).map(|i| (i as f32) * 0.01).collect();

    let a = trainer.upload(&a_data);
    let b = trainer.upload(&b_data);
    let c = trainer.zeros((m * n) as usize);

    trainer.matmul_forward(&a, &b, &c, m, k, n);
    let result = trainer.download(&c);

    // Compute expected on CPU
    let mut expected = vec![0.0f32; (m * n) as usize];
    for i in 0..m as usize {
        for j in 0..n as usize {
            for kk in 0..k as usize {
                expected[i * n as usize + j] +=
                    a_data[i * k as usize + kk] * b_data[kk * n as usize + j];
            }
        }
    }

    for i in 0..(m * n) as usize {
        let err = (result[i] - expected[i]).abs();
        assert!(err < 1e-3, "Mismatch at {}: gpu={} cpu={} err={}", i, result[i], expected[i], err);
    }
}

/// FALSIFY-WGSL-GEMM-002: Backward GEMM correctness
#[test]
fn test_wgpu_matmul_backward() {
    let trainer = match WgpuTrainer::new() {
        Ok(t) => t,
        Err(_) => return,
    };

    let m = 4u32;
    let k = 3u32;
    let n = 5u32;

    let a_data: Vec<f32> = (0..m * k).map(|i| (i as f32) * 0.1).collect();
    let b_data: Vec<f32> = (0..k * n).map(|i| (i as f32) * 0.01).collect();
    let gc_data: Vec<f32> = (0..m * n).map(|i| (i as f32) * 0.05).collect();

    let a = trainer.upload(&a_data);
    let b = trainer.upload(&b_data);
    let gc = trainer.upload(&gc_data);
    let ga = trainer.zeros((m * k) as usize);
    let gb = trainer.zeros((k * n) as usize);

    trainer.matmul_backward(&a, &b, &gc, &ga, &gb, m, k, n);
    let ga_result = trainer.download(&ga);
    let gb_result = trainer.download(&gb);

    // CPU reference: grad_a = grad_c @ B^T
    let mut ga_expected = vec![0.0f32; (m * k) as usize];
    for i in 0..m as usize {
        for j in 0..k as usize {
            for nn in 0..n as usize {
                ga_expected[i * k as usize + j] +=
                    gc_data[i * n as usize + nn] * b_data[j * n as usize + nn];
            }
        }
    }

    // CPU reference: grad_b = A^T @ grad_c
    let mut gb_expected = vec![0.0f32; (k * n) as usize];
    for i in 0..k as usize {
        for j in 0..n as usize {
            for mm in 0..m as usize {
                gb_expected[i * n as usize + j] +=
                    a_data[mm * k as usize + i] * gc_data[mm * n as usize + j];
            }
        }
    }

    for i in 0..(m * k) as usize {
        let err = (ga_result[i] - ga_expected[i]).abs();
        assert!(err < 1e-3, "grad_a[{}]: gpu={} cpu={}", i, ga_result[i], ga_expected[i]);
    }
    for i in 0..(k * n) as usize {
        let err = (gb_result[i] - gb_expected[i]).abs();
        assert!(err < 1e-3, "grad_b[{}]: gpu={} cpu={}", i, gb_result[i], gb_expected[i]);
    }
}

/// AdamW updates params in correct direction
#[test]
fn test_wgpu_adamw_step() {
    let mut trainer = match WgpuTrainer::new() {
        Ok(t) => t,
        Err(_) => return,
    };

    let params = trainer.upload(&[1.0, 2.0, 3.0, 4.0]);
    let grads = trainer.upload(&[0.1, 0.2, 0.3, 0.4]);
    let m_state = trainer.zeros(4);
    let v_state = trainer.zeros(4);

    trainer.adamw_step(&params, &grads, &m_state, &v_state, 0.001, 0.9, 0.999, 1e-8, 0.01);

    let result = trainer.download(&params);
    // Params should have decreased (moved toward lower loss)
    assert!(result[0] < 1.0, "param[0] should decrease: {}", result[0]);
    assert!(result[3] < 4.0, "param[3] should decrease: {}", result[3]);
}

/// Test GEMM with LoRA-sized dimensions (M=64, K=512, N=3584)
#[test]
fn test_wgpu_matmul_lora_dims() {
    let trainer = match WgpuTrainer::new() {
        Ok(t) => t,
        Err(_) => return,
    };

    let m = 64u32; // rank
    let k = 512u32; // seq_len
    let n = 128u32; // smaller out_dim for fast test

    let a_data: Vec<f32> = (0..m * k).map(|i| ((i as f32) * 0.01).sin()).collect();
    let b_data: Vec<f32> = (0..k * n).map(|i| ((i as f32) * 0.007).cos()).collect();

    let a = trainer.upload(&a_data);
    let b = trainer.upload(&b_data);
    let c = trainer.zeros((m * n) as usize);

    trainer.matmul_forward(&a, &b, &c, m, k, n);
    let result = trainer.download(&c);

    let mut expected = vec![0.0f32; (m * n) as usize];
    for i in 0..m as usize {
        for j in 0..n as usize {
            for kk in 0..k as usize {
                expected[i * n as usize + j] +=
                    a_data[i * k as usize + kk] * b_data[kk * n as usize + j];
            }
        }
    }

    let mut max_err = 0.0f32;
    let mut result_norm = 0.0f32;
    for i in 0..(m * n) as usize {
        let err = (result[i] - expected[i]).abs();
        max_err = max_err.max(err);
        result_norm += result[i] * result[i];
    }
    result_norm = result_norm.sqrt();

    eprintln!("LoRA GEMM: result_norm={result_norm:.4}, max_err={max_err:.6}");
    assert!(result_norm > 0.0, "GEMM result is all zeros!");
    assert!(max_err < 1.0, "GEMM error too large: {max_err}");
}

/// Test backward with from_device (shared device)
#[test]
fn test_wgpu_matmul_backward_shared_device() {
    let instance = wgpu::Instance::new(&wgpu::InstanceDescriptor::default());
    let adapter = match trueno::backends::gpu::runtime::block_on(
        instance.request_adapter(&wgpu::RequestAdapterOptions::default()),
    ) {
        Ok(a) => a,
        Err(_) => return,
    };
    let (device, queue) = match trueno::backends::gpu::runtime::block_on(
        adapter.request_device(&wgpu::DeviceDescriptor::default()),
    ) {
        Ok(dq) => dq,
        Err(_) => return,
    };

    let trainer = match WgpuTrainer::from_device(device, queue) {
        Ok(t) => t,
        Err(_) => return,
    };

    let m = 64u32;
    let k = 32u32;
    let n = 128u32;
    let a_data: Vec<f32> = (0..m * k).map(|i| ((i as f32) * 0.013).sin()).collect();
    let b_data: Vec<f32> = (0..k * n).map(|i| ((i as f32) * 0.009).cos()).collect();
    let gc_data: Vec<f32> = (0..m * n).map(|i| ((i as f32) * 0.011).sin()).collect();

    let a = trainer.upload(&a_data);
    let b = trainer.upload(&b_data);
    let gc = trainer.upload(&gc_data);
    let ga = trainer.zeros((m * k) as usize);
    let gb = trainer.zeros((k * n) as usize);

    trainer.matmul_backward(&a, &b, &gc, &ga, &gb, m, k, n);
    let gb_result = trainer.download(&gb);

    let gb_norm: f32 = gb_result.iter().map(|x| x * x).sum::<f32>().sqrt();
    eprintln!("Shared device backward: gb_norm={gb_norm:.6}");
    assert!(gb_norm > 0.0, "grad_b should be non-zero");
}