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
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


/// Test transformer_layer_batched positions mismatch
#[test]
#[serial]
fn test_cov030_transformer_layer_batched_positions_mismatch() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    // Init workspace for batch_size=4
    executor.init_workspace(512, 256).expect("init workspace");
    executor
        .init_batched_workspace(512, 256, 4)
        .expect("init batched");

    let input_data = vec![0.1f32; 512 * 4];
    let input = GpuBuffer::from_host(executor.context(), &input_data).expect("input");
    let layer_weights = test_zeroed_layer_weights();

    // Positions length != m
    let result = executor.transformer_layer_batched(
        &input,
        0,
        &layer_weights,
        4,
        &[0, 1], // only 2 positions but m=4
        512,
        256,
        1e-5,
    );
    assert!(result.is_err(), "Should error when positions.len() != m");
}

// =============================================================================
// COV-031: Additional Activation & Attention Coverage Tests
// Target: Improve activations.rs and attention.rs coverage
// =============================================================================

/// Test rope_into basic operation
#[test]
#[serial]
fn test_cov031_rope_into_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 8u32;
    let head_dim = 64u32;
    let n = num_heads * head_dim;
    let theta = 10000.0f32;
    let position = 5u32;

    // Create input/output buffers
    let input_data = vec![1.0f32; n as usize];
    let input = GpuBuffer::from_host(executor.context(), &input_data).expect("input");
    let output = GpuBuffer::new(executor.context(), n as usize).expect("output");

    let result = executor.rope_into(&input, &output, position, num_heads, head_dim, theta);
    assert!(
        result.is_ok(),
        "rope_into should succeed: {:?}",
        result.err()
    );
}

/// Test rope_into with different positions
#[test]
#[serial]
fn test_cov031_rope_into_positions() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 8u32;
    let head_dim = 64u32;
    let n = num_heads * head_dim;
    let theta = 10000.0f32;

    let input_data = vec![1.0f32; n as usize];
    let input = GpuBuffer::from_host(executor.context(), &input_data).expect("input");
    let output = GpuBuffer::new(executor.context(), n as usize).expect("output");

    // Test various positions
    for position in [0u32, 1, 10, 100, 1000] {
        let result = executor.rope_into(&input, &output, position, num_heads, head_dim, theta);
        assert!(
            result.is_ok(),
            "rope_into at position {} failed: {:?}",
            position,
            result.err()
        );
    }
}

/// Test rope_indirect_into basic operation
#[test]
#[serial]
fn test_cov031_rope_indirect_into_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 8u32;
    let head_dim = 64u32;
    let n = num_heads * head_dim;
    let theta = 10000.0f32;

    let input_data = vec![1.0f32; n as usize];
    let input = GpuBuffer::from_host(executor.context(), &input_data).expect("input");
    let output = GpuBuffer::new(executor.context(), n as usize).expect("output");

    // Position in device buffer (for CUDA graph compatibility)
    let position_buf = GpuBuffer::from_host(executor.context(), &[5u32]).expect("position buf");

    let result =
        executor.rope_indirect_into(&input, &output, &position_buf, num_heads, head_dim, theta);
    assert!(
        result.is_ok(),
        "rope_indirect_into should succeed: {:?}",
        result.err()
    );
}

/// Test rope_neox_into basic operation
#[test]
#[serial]
fn test_cov031_rope_neox_into_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 8u32;
    let head_dim = 64u32;
    let n = num_heads * head_dim;
    let theta = 10000.0f32;
    let position = 5u32;

    let input_data = vec![1.0f32; n as usize];
    let input = GpuBuffer::from_host(executor.context(), &input_data).expect("input");
    let output = GpuBuffer::new(executor.context(), n as usize).expect("output");

    let result = executor.rope_neox_into(&input, &output, position, num_heads, head_dim, theta);
    assert!(
        result.is_ok(),
        "rope_neox_into should succeed: {:?}",
        result.err()
    );
}

/// Test rope_neox_indirect_into basic operation
#[test]
#[serial]
fn test_cov031_rope_neox_indirect_into_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 8u32;
    let head_dim = 64u32;
    let n = num_heads * head_dim;
    let theta = 10000.0f32;

    let input_data = vec![1.0f32; n as usize];
    let input = GpuBuffer::from_host(executor.context(), &input_data).expect("input");
    let output = GpuBuffer::new(executor.context(), n as usize).expect("output");
    let position_buf = GpuBuffer::from_host(executor.context(), &[10u32]).expect("position buf");

    let result = executor.rope_neox_indirect_into(
        &input,
        &output,
        &position_buf,
        num_heads,
        head_dim,
        theta,
    );
    assert!(
        result.is_ok(),
        "rope_neox_indirect_into should succeed: {:?}",
        result.err()
    );
}

/// Test fused_qkv_into basic operation
#[test]
#[serial]
fn test_cov031_fused_qkv_into_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    // Use small dimensions to avoid memory issues
    // fused_qkv_into expects: x[hidden_size], w_q[hidden_size, hidden_size],
    // w_k[hidden_size, kv_dim], w_v[hidden_size, kv_dim]
    // out_q[hidden_size], out_k[kv_dim], out_v[kv_dim]
    let hidden_dim = 64u32;
    let kv_dim = 32u32; // GQA with fewer KV heads

    // Create weight matrices as f32 GpuBuffers
    let w_q_data = vec![0.01f32; (hidden_dim * hidden_dim) as usize]; // Q output is hidden_dim
    let w_k_data = vec![0.01f32; (hidden_dim * kv_dim) as usize];
    let w_v_data = vec![0.01f32; (hidden_dim * kv_dim) as usize];

    let w_q = GpuBuffer::from_host(executor.context(), &w_q_data).expect("w_q");
    let w_k = GpuBuffer::from_host(executor.context(), &w_k_data).expect("w_k");
    let w_v = GpuBuffer::from_host(executor.context(), &w_v_data).expect("w_v");

    let input_data = vec![0.1f32; hidden_dim as usize];
    let input = GpuBuffer::from_host(executor.context(), &input_data).expect("input");

    // Q output has hidden_dim elements, K/V have kv_dim elements
    let q_out = GpuBuffer::new(executor.context(), hidden_dim as usize).expect("q_out");
    let k_out = GpuBuffer::new(executor.context(), kv_dim as usize).expect("k_out");
    let v_out = GpuBuffer::new(executor.context(), kv_dim as usize).expect("v_out");

    let result = executor.fused_qkv_into(
        &input, &w_q, &w_k, &w_v, &q_out, &k_out, &v_out, hidden_dim, kv_dim,
    );
    assert!(
        result.is_ok(),
        "fused_qkv_into should succeed: {:?}",
        result.err()
    );

    // Synchronize to catch any kernel errors before test ends
    executor.synchronize().expect("sync");
}

/// Test fused_gate_up_into basic operation
#[test]
#[serial]
fn test_cov031_fused_gate_up_into_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    // Use smaller dimensions to avoid memory issues
    let hidden_dim = 64u32;
    let intermediate_dim = 128u32;

    // Create weight matrices as f32 GpuBuffers
    let w_gate_data = vec![0.01f32; (hidden_dim * intermediate_dim) as usize];
    let w_up_data = vec![0.01f32; (hidden_dim * intermediate_dim) as usize];

    let w_gate = GpuBuffer::from_host(executor.context(), &w_gate_data).expect("w_gate");
    let w_up = GpuBuffer::from_host(executor.context(), &w_up_data).expect("w_up");

    let input_data = vec![0.1f32; hidden_dim as usize];
    let input = GpuBuffer::from_host(executor.context(), &input_data).expect("input");

    let output = GpuBuffer::new(executor.context(), intermediate_dim as usize).expect("output");

    let result = executor.fused_gate_up_into(
        &input,
        &w_gate,
        &w_up,
        &output,
        hidden_dim,
        intermediate_dim,
    );
    assert!(
        result.is_ok(),
        "fused_gate_up_into should succeed: {:?}",
        result.err()
    );

    // Synchronize to catch any kernel errors before test ends
    executor.synchronize().expect("sync");
}

/// Test incremental_attention_into basic operation
#[test]
#[serial]
fn test_cov031_incremental_attention_into_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 4usize;
    let num_kv_heads = 4usize;
    let head_dim = 8usize;
    let q_dim = num_heads * head_dim;
    let kv_dim = num_kv_heads * head_dim;

    // Initialize KV cache first (required for incremental attention)
    executor
        .init_kv_cache_gpu(1, num_heads, num_kv_heads, head_dim, 16)
        .expect("init kv");

    let q_data = vec![0.1f32; q_dim];
    let k_data = vec![0.1f32; kv_dim];
    let v_data = vec![0.1f32; kv_dim];

    let q_buf = GpuBuffer::from_host(executor.context(), &q_data).expect("q_buf");
    let k_buf = GpuBuffer::from_host(executor.context(), &k_data).expect("k_buf");
    let v_buf = GpuBuffer::from_host(executor.context(), &v_data).expect("v_buf");
    let out_buf = GpuBuffer::new(executor.context(), q_dim).expect("out_buf");

    let result = executor.incremental_attention_into(0, &q_buf, &k_buf, &v_buf, &out_buf);
    assert!(
        result.is_ok(),
        "incremental_attention_into should succeed: {:?}",
        result.err()
    );
}

/// Test batched_incremental_attention_into with batch_size=2
#[test]
#[serial]
fn test_cov031_batched_incremental_attention_into() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 4usize;
    let num_kv_heads = 4usize;
    let head_dim = 8usize;
    let batch_size = 2usize;
    let q_dim = num_heads * head_dim;
    let kv_dim = num_kv_heads * head_dim;

    // Initialize KV cache first (required for batched attention)
    executor
        .init_kv_cache_gpu(1, num_heads, num_kv_heads, head_dim, 16)
        .expect("init kv");
    executor
        .init_batched_kv_cache_gpu(1, batch_size)
        .expect("init batched kv");

    let q_data = vec![0.1f32; q_dim * batch_size];
    let k_data = vec![0.1f32; kv_dim * batch_size];
    let v_data = vec![0.1f32; kv_dim * batch_size];

    let q_buf = GpuBuffer::from_host(executor.context(), &q_data).expect("q_buf");
    let k_buf = GpuBuffer::from_host(executor.context(), &k_data).expect("k_buf");
    let v_buf = GpuBuffer::from_host(executor.context(), &v_data).expect("v_buf");
    let out_buf = GpuBuffer::new(executor.context(), q_dim * batch_size).expect("out_buf");

    let positions = vec![0u32; batch_size];

    let result = executor.batched_incremental_attention_into(
        0, &q_buf, &k_buf, &v_buf, &out_buf, batch_size, &positions,
    );
    assert!(
        result.is_ok(),
        "batched_incremental_attention_into should succeed: {:?}",
        result.err()
    );
}

/// Test flash_decoding_attention_into without init
#[test]
#[serial]
fn test_cov031_flash_decoding_not_init() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let seq_len = 16usize;
    let head_dim = 64usize;
    let n = seq_len * head_dim;

    let q_data = vec![0.1f32; n];
    let k_data = vec![0.1f32; n];
    let v_data = vec![0.1f32; n];

    let q_buf = GpuBuffer::from_host(executor.context(), &q_data).expect("q_buf");
    let k_buf = GpuBuffer::from_host(executor.context(), &k_data).expect("k_buf");
    let v_buf = GpuBuffer::from_host(executor.context(), &v_data).expect("v_buf");
    let out_buf = GpuBuffer::new(executor.context(), n).expect("out_buf");

    // Without init_flash_decoding, should return error
    let positions = vec![0u32; 1];
    let result =
        executor.flash_decoding_attention_into(0, &q_buf, &k_buf, &v_buf, &out_buf, 1, &positions);
    assert!(result.is_err(), "flash_decoding should error without init");
}

/// Test init_flash_decoding and flash_decoding_attention_into
#[test]
#[serial]
fn test_cov031_flash_decoding_with_init() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let num_heads = 8usize;
    let head_dim = 64usize;
    let max_seq_len = 128usize;
    let batch_size = 1usize;

    // Initialize flash decoding (num_heads, head_dim, max_seq_len, batch_size)
    let init_result = executor.init_flash_decoding(num_heads, head_dim, max_seq_len, batch_size);
    assert!(
        init_result.is_ok(),
        "init_flash_decoding should succeed: {:?}",
        init_result.err()
    );

    // Now try flash decoding attention
    let q_dim = num_heads * head_dim;

    let q_data = vec![0.1f32; q_dim];
    let k_data = vec![0.1f32; q_dim];
    let v_data = vec![0.1f32; q_dim];

    let q_buf = GpuBuffer::from_host(executor.context(), &q_data).expect("q_buf");
    let k_buf = GpuBuffer::from_host(executor.context(), &k_data).expect("k_buf");
    let v_buf = GpuBuffer::from_host(executor.context(), &v_data).expect("v_buf");
    let out_buf = GpuBuffer::new(executor.context(), q_dim).expect("out_buf");

    let positions = vec![0u32; batch_size];
    let result = executor
        .flash_decoding_attention_into(0, &q_buf, &k_buf, &v_buf, &out_buf, batch_size, &positions);
    // Note: flash decoding may fail if KV cache not initialized, but at least we cover init path
    let _ = result;
}