realizar 0.8.5

Pure Rust ML inference engine built from scratch - model serving for GGUF and safetensors
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


#[test]
#[serial]
fn test_cov007_elementwise_mul_gpu_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 128u32;
    let a_data: Vec<f32> = (0..n).map(|i| i as f32).collect();
    let b_data: Vec<f32> = (0..n).map(|_| 2.0f32).collect();

    let a = GpuBuffer::from_host(&executor.context, &a_data).expect("a buffer");
    let b = GpuBuffer::from_host(&executor.context, &b_data).expect("b buffer");

    let result = executor.elementwise_mul_gpu(&a, &b, n);
    assert!(
        result.is_ok(),
        "elementwise_mul_gpu failed: {:?}",
        result.err()
    );

    let output_buffer = result.unwrap();
    executor.stream.synchronize().expect("sync");

    let mut output = vec![0.0f32; n as usize];
    output_buffer.copy_to_host(&mut output).expect("copy");

    // output[i] = a[i] * b[i] = i * 2 = 2i
    for (idx, &val) in output.iter().enumerate() {
        let expected = (idx as f32) * 2.0;
        assert!(
            (val - expected).abs() < 1e-4,
            "mul mismatch at {}: {} vs {}",
            idx,
            val,
            expected
        );
    }
}

#[test]
#[serial]
fn test_cov007_elementwise_mul_gpu_large() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 4096u32;
    let a = GpuBuffer::from_host(&executor.context, &vec![3.0f32; n as usize]).expect("a");
    let b = GpuBuffer::from_host(&executor.context, &vec![4.0f32; n as usize]).expect("b");

    let result = executor.elementwise_mul_gpu(&a, &b, n);
    assert!(
        result.is_ok(),
        "elementwise_mul_gpu large failed: {:?}",
        result.err()
    );

    let output_buffer = result.unwrap();
    executor.stream.synchronize().expect("sync");

    let mut output = vec![0.0f32; n as usize];
    output_buffer.copy_to_host(&mut output).expect("copy");

    // All should be 12.0
    assert!(
        output.iter().all(|&x| (x - 12.0).abs() < 1e-4),
        "All outputs should be 12.0"
    );
}

#[test]
#[serial]
fn test_cov007_silu_host_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 64usize;
    let input: Vec<f32> = (0..n).map(|i| (i as f32 - 32.0) / 16.0).collect();
    let mut output = vec![0.0f32; n];

    let result = executor.silu_host(&input, &mut output);
    assert!(result.is_ok(), "silu_host failed: {:?}", result.err());

    // SiLU(0) should be near 0
    let mid_idx = 32;
    assert!(output[mid_idx].abs() < 0.1, "SiLU(0) should be near 0");
}

#[test]
#[serial]
fn test_cov007_silu_host_large() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 512usize;
    let input = vec![1.0f32; n];
    let mut output = vec![0.0f32; n];

    let result = executor.silu_host(&input, &mut output);
    assert!(result.is_ok(), "silu_host large failed: {:?}", result.err());

    // SiLU(1) = 1 * sigmoid(1) ≈ 0.731
    assert!(
        output[0] > 0.7 && output[0] < 0.8,
        "SiLU(1) should be ~0.731, got {}",
        output[0]
    );
}

#[test]
#[serial]
fn test_cov007_gelu_host_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 64usize;
    let input: Vec<f32> = (0..n).map(|i| (i as f32 - 32.0) / 16.0).collect();
    let mut output = vec![0.0f32; n];

    let result = executor.gelu_host(&input, &mut output);
    assert!(result.is_ok(), "gelu_host failed: {:?}", result.err());

    // GELU(0) should be near 0
    let mid_idx = 32;
    assert!(output[mid_idx].abs() < 0.1, "GELU(0) should be near 0");
}

#[test]
#[serial]
fn test_cov007_gelu_host_positive() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 64usize;
    let input = vec![2.0f32; n]; // All 2.0
    let mut output = vec![0.0f32; n];

    let result = executor.gelu_host(&input, &mut output);
    assert!(
        result.is_ok(),
        "gelu_host positive failed: {:?}",
        result.err()
    );

    // GELU(2) should be close to 2 (slightly less)
    assert!(
        output[0] > 1.9 && output[0] < 2.1,
        "GELU(2) should be ~2.0, got {}",
        output[0]
    );
}

#[test]
#[serial]
fn test_cov007_elementwise_mul_host_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 64usize;
    let a: Vec<f32> = (0..n).map(|i| i as f32).collect();
    let b: Vec<f32> = (0..n).map(|i| (n - i) as f32).collect();
    let mut output = vec![0.0f32; n];

    let result = executor.elementwise_mul_host(&a, &b, &mut output);
    assert!(
        result.is_ok(),
        "elementwise_mul_host failed: {:?}",
        result.err()
    );

    // output[i] = i * (n - i)
    for (idx, &val) in output.iter().enumerate() {
        let expected = (idx as f32) * ((n - idx) as f32);
        assert!(
            (val - expected).abs() < 1e-4,
            "mul_host mismatch at {}: {} vs {}",
            idx,
            val,
            expected
        );
    }
}

#[test]
#[serial]
fn test_cov007_fused_swiglu_host_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 64usize;
    let gate = vec![1.0f32; n];
    let up = vec![2.0f32; n];
    let mut output = vec![0.0f32; n];

    let result = executor.fused_swiglu_host(&gate, &up, &mut output);
    assert!(
        result.is_ok(),
        "fused_swiglu_host failed: {:?}",
        result.err()
    );

    // SwiGLU(gate, up) = silu(gate) * up = silu(1) * 2 ≈ 0.731 * 2 ≈ 1.462
    assert!(
        output[0] > 1.4 && output[0] < 1.6,
        "SwiGLU(1,2) should be ~1.46, got {}",
        output[0]
    );
}

#[test]
#[serial]
fn test_cov007_fused_swiglu_host_large() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 256usize;
    let gate: Vec<f32> = (0..n).map(|i| (i as f32) / 128.0).collect();
    let up = vec![1.0f32; n];
    let mut output = vec![0.0f32; n];

    let result = executor.fused_swiglu_host(&gate, &up, &mut output);
    assert!(
        result.is_ok(),
        "fused_swiglu_host large failed: {:?}",
        result.err()
    );
}

#[test]
#[serial]
fn test_cov007_add_residual_gpu_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 128u32;

    // Output starts with values, input is what to add
    let output_data: Vec<f32> = (0..n).map(|i| i as f32).collect();
    let input_data: Vec<f32> = (0..n).map(|_| 10.0f32).collect();

    let output_buf = GpuBuffer::from_host(&executor.context, &output_data).expect("output buffer");
    let input_buf = GpuBuffer::from_host(&executor.context, &input_data).expect("input buffer");

    let result = executor.add_residual_gpu(&output_buf, &input_buf, n);
    assert!(
        result.is_ok(),
        "add_residual_gpu failed: {:?}",
        result.err()
    );

    executor.stream.synchronize().expect("sync");

    let mut output = vec![0.0f32; n as usize];
    output_buf.copy_to_host(&mut output).expect("copy");

    // output[i] += 10, so output[i] = i + 10
    for (idx, &val) in output.iter().enumerate() {
        let expected = idx as f32 + 10.0;
        assert!(
            (val - expected).abs() < 1e-4,
            "add_residual mismatch at {}: {} vs {}",
            idx,
            val,
            expected
        );
    }
}

#[test]
#[serial]
fn test_cov007_add_residual_gpu_large() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 4096u32;

    let output_buf =
        GpuBuffer::from_host(&executor.context, &vec![5.0f32; n as usize]).expect("output");
    let input_buf =
        GpuBuffer::from_host(&executor.context, &vec![3.0f32; n as usize]).expect("input");

    let result = executor.add_residual_gpu(&output_buf, &input_buf, n);
    assert!(
        result.is_ok(),
        "add_residual_gpu large failed: {:?}",
        result.err()
    );

    executor.stream.synchronize().expect("sync");

    let mut output = vec![0.0f32; n as usize];
    output_buf.copy_to_host(&mut output).expect("copy");

    // All should be 8.0 (5 + 3)
    assert!(
        output.iter().all(|&x| (x - 8.0).abs() < 1e-4),
        "All outputs should be 8.0"
    );
}

#[test]
#[serial]
fn test_cov007_fused_swiglu_gpu_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 256u32;
    let gate_data = vec![1.0f32; n as usize];
    let up_data = vec![2.0f32; n as usize];

    let gate = GpuBuffer::from_host(&executor.context, &gate_data).expect("gate buffer");
    let up = GpuBuffer::from_host(&executor.context, &up_data).expect("up buffer");

    let result = executor.fused_swiglu_gpu(&gate, &up, n);
    assert!(
        result.is_ok(),
        "fused_swiglu_gpu failed: {:?}",
        result.err()
    );

    let output_buffer = result.unwrap();
    executor.stream.synchronize().expect("sync");

    let mut output = vec![0.0f32; n as usize];
    output_buffer.copy_to_host(&mut output).expect("copy");

    // SwiGLU(1,2) = silu(1) * 2 ≈ 1.46
    assert!(
        output[0] > 1.4 && output[0] < 1.6,
        "SwiGLU(1,2) should be ~1.46, got {}",
        output[0]
    );
}

#[test]
#[serial]
fn test_cov007_fused_swiglu_gpu_large() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 2048u32;

    let gate = GpuBuffer::from_host(&executor.context, &vec![0.5f32; n as usize]).expect("gate");
    let up = GpuBuffer::from_host(&executor.context, &vec![1.0f32; n as usize]).expect("up");

    let result = executor.fused_swiglu_gpu(&gate, &up, n);
    assert!(
        result.is_ok(),
        "fused_swiglu_gpu large failed: {:?}",
        result.err()
    );
}