aprender-serve 0.65.2

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
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


// ============================================================================
// forward_with_cache: Q6K lm_head (different from Q4K lm_head path)
// ============================================================================

#[test]
fn test_fwc_q6k_lm_head_with_trace() {
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::set_var("REALIZE_TRACE", "1");
    }
    let mut apr = build_apr_with_q4k_fused(32, 64, 4, 4, 16);
    apr.lm_head_weight_q6k = Some(q6k_bytes(16, 32));
    let mut cache = AprKVCache::new(&apr.config);

    let result = apr.forward_with_cache(1, &mut cache, 0);
    assert!(
        result.is_ok(),
        "Q6K lm_head + trace: {}",
        result.unwrap_err()
    );
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::remove_var("REALIZE_TRACE");
    }
}

// ============================================================================
// forward_with_cache: Q4K lm_head with trace
// ============================================================================

#[test]
fn test_fwc_q4k_lm_head_with_trace() {
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::set_var("REALIZE_TRACE", "1");
    }
    let mut apr = build_apr_with_q4k_fused(32, 64, 4, 4, 16);
    apr.lm_head_weight_q4k = Some(q4k_bytes(16, 32));
    let mut cache = AprKVCache::new(&apr.config);

    let result = apr.forward_with_cache(1, &mut cache, 0);
    assert!(
        result.is_ok(),
        "Q4K lm_head + trace: {}",
        result.unwrap_err()
    );
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::remove_var("REALIZE_TRACE");
    }
}

// ============================================================================
// forward_with_cache: Q6K variants with trace (covers Q6K trace blocks)
// ============================================================================

#[test]
fn test_fwc_q6k_variants_with_trace() {
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::set_var("REALIZE_TRACE", "1");
    }
    let apr = build_apr_with_q6k_variants(32, 64, 16);
    let mut cache = AprKVCache::new(&apr.config);

    let r1 = apr.forward_with_cache(1, &mut cache, 0);
    assert!(r1.is_ok(), "Q6K variants + trace: {}", r1.unwrap_err());
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::remove_var("REALIZE_TRACE");
    }
}

// ============================================================================
// forward_with_cache: partial Q4K (some fields None) + trace
// Covers the F32 fallback trace blocks INSIDE the `if !force_f32` branch
// (e.g., "attn_output using F32 fallback (slow!)")
// ============================================================================

#[test]
fn test_fwc_partial_q4k_f32_fallback_with_trace() {
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::set_var("REALIZE_TRACE", "1");
    }

    let hidden = 32;
    let intermediate = 64;
    let vocab = 16;
    let heads = 4;
    let kv_heads = 4;
    let head_dim = hidden / heads;
    let kv_size = kv_heads * head_dim;

    let layer = AprTransformerLayer {
        attn_norm_weight: vec![1.0; hidden],
        attn_norm_bias: None,
        qkv_weight: vec![0.001; (hidden + 2 * kv_size) * hidden],
        qkv_bias: None,
        attn_output_weight: vec![0.001; hidden * hidden],
        attn_output_bias: None,
        ffn_gate_weight: Some(vec![0.001; intermediate * hidden]),
        ffn_gate_bias: None,
        ffn_up_weight: vec![0.001; intermediate * hidden],
        ffn_up_bias: None,
        ffn_down_weight: vec![0.001; hidden * intermediate],
        ffn_down_bias: None,
        ffn_norm_weight: Some(vec![1.0; hidden]),
        ffn_norm_bias: None,
        attn_q_norm_weight: None,
        attn_k_norm_weight: None,
        linear_attn_z_weight: None,
        linear_attn_b_weight: None,
        linear_attn_a_weight: None,
        linear_attn_conv1d_weight: None,
        linear_attn_a_log: None,
        linear_attn_dt_bias: None,
        linear_attn_norm_weight: None,
        moe_gate_weight: None,
        moe_expert_gate_up: None,
        moe_expert_down: None,
        moe_shared_gate: None,
        moe_shared_up: None,
        moe_shared_down: None,
        moe_shared_expert_gate_weight: None,
    };

    // Partial Q4K: only Q and K populated, everything else None
    // This forces F32 fallback for V, attn_output, gate, up, down
    let q4k = Q4KLayerWeights {
        qkv_weight: None,
        attn_q_weight: Some(q4k_bytes(hidden, hidden)),
        attn_k_weight: Some(q4k_bytes(kv_size, hidden)),
        attn_v_weight: None,      // F32 fallback
        attn_v_weight_q6k: None,  // no Q6K either
        attn_output_weight: None, // F32 fallback
        ffn_gate_weight: None,    // F32 fallback
        ffn_up_weight: None,      // F32 fallback
        ffn_down_weight: None,    // F32 fallback
        ffn_down_weight_q6k: None,
        ffn_up_weight_q6k: None,
    };

    let apr = AprTransformer {
        config: AprTransformerConfig {
            architecture: "llama".to_string(),
            hidden_dim: hidden,
            num_layers: 1,
            num_heads: heads,
            num_kv_heads: kv_heads,
            vocab_size: vocab,
            intermediate_dim: intermediate,
            context_length: 2048,
            rope_theta: 10000.0,
            eps: 1e-5,
            eos_token_id: None,
        ..Default::default()
        },
        token_embedding: vec![0.01; vocab * hidden],
        layers: vec![layer],
        output_norm_weight: vec![1.0; hidden],
        output_norm_bias: None,
        lm_head_weight: vec![0.01; vocab * hidden],
        lm_head_bias: None,
        lm_head_tied: false,
        q4k_layers: Some(vec![q4k]),
        lm_head_weight_q6k: None,
        lm_head_weight_q4k: None,
    };

    let mut cache = AprKVCache::new(&apr.config);
    let r1 = apr.forward_with_cache(1, &mut cache, 0);
    assert!(r1.is_ok(), "Partial Q4K + trace: {}", r1.unwrap_err());
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::remove_var("REALIZE_TRACE");
    }
}

// ============================================================================
// forward (batch): out-of-vocab token + debug traces
// ============================================================================

#[test]
fn test_forward_batch_oov_token_with_debug() {
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::set_var("REALIZE_DEBUG", "1");
    }
    let apr = build_apr_with_q4k_fused(32, 64, 4, 4, 16);
    // Token 999 is way out of vocab (vocab=16)
    let result = apr.forward(&[999]);
    assert!(result.is_ok(), "OOV batch: {}", result.unwrap_err());
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::remove_var("REALIZE_DEBUG");
    }
}

// ============================================================================
// forward_with_cache: GELU model (no gate) F32-only + trace
// Covers GELU branches in forward_with_cache with trace enabled
// ============================================================================

#[test]
fn test_fwc_gelu_f32_no_q4k_with_trace() {
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::set_var("REALIZE_TRACE", "1");
    }

    let hidden = 32;
    let intermediate = 64;
    let vocab = 16;

    let layer = AprTransformerLayer {
        attn_norm_weight: vec![1.0; hidden],
        attn_norm_bias: Some(vec![0.0; hidden]),
        qkv_weight: vec![0.001; 3 * hidden * hidden],
        qkv_bias: Some(vec![0.0; 3 * hidden]),
        attn_output_weight: vec![0.001; hidden * hidden],
        attn_output_bias: Some(vec![0.0; hidden]),
        ffn_gate_weight: None, // GELU
        ffn_gate_bias: None,
        ffn_up_weight: vec![0.001; intermediate * hidden],
        ffn_up_bias: Some(vec![0.0; intermediate]),
        ffn_down_weight: vec![0.001; hidden * intermediate],
        ffn_down_bias: Some(vec![0.0; hidden]),
        ffn_norm_weight: Some(vec![1.0; hidden]),
        ffn_norm_bias: Some(vec![0.0; hidden]),
        attn_q_norm_weight: None,
        attn_k_norm_weight: None,
        linear_attn_z_weight: None,
        linear_attn_b_weight: None,
        linear_attn_a_weight: None,
        linear_attn_conv1d_weight: None,
        linear_attn_a_log: None,
        linear_attn_dt_bias: None,
        linear_attn_norm_weight: None,
        moe_gate_weight: None,
        moe_expert_gate_up: None,
        moe_expert_down: None,
        moe_shared_gate: None,
        moe_shared_up: None,
        moe_shared_down: None,
        moe_shared_expert_gate_weight: None,
    };

    let apr = AprTransformer {
        config: AprTransformerConfig {
            architecture: "phi2".to_string(),
            hidden_dim: hidden,
            num_layers: 1,
            num_heads: 4,
            num_kv_heads: 4,
            vocab_size: vocab,
            intermediate_dim: intermediate,
            context_length: 2048,
            rope_theta: 10000.0,
            eps: 1e-5,
            eos_token_id: None,
        ..Default::default()
        },
        token_embedding: vec![0.01; vocab * hidden],
        layers: vec![layer],
        output_norm_weight: vec![1.0; hidden],
        output_norm_bias: Some(vec![0.0; hidden]),
        lm_head_weight: vec![0.01; vocab * hidden],
        lm_head_bias: Some(vec![0.0; vocab]),
        lm_head_tied: false,
        q4k_layers: None,
        lm_head_weight_q6k: None,
        lm_head_weight_q4k: None,
    };

    let mut cache = AprKVCache::new(&apr.config);
    let result = apr.forward_with_cache(1, &mut cache, 0);
    assert!(result.is_ok(), "GELU F32 + trace: {}", result.unwrap_err());
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::remove_var("REALIZE_TRACE");
    }
}

// ============================================================================
// forward (batch): GELU + trace (covers batch GELU trace paths)
// ============================================================================

#[test]
fn test_forward_batch_gelu_with_trace() {
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::set_var("REALIZE_TRACE", "1");
    }

    let hidden = 32;
    let intermediate = 64;
    let vocab = 16;

    let layer = AprTransformerLayer {
        attn_norm_weight: vec![1.0; hidden],
        attn_norm_bias: Some(vec![0.0; hidden]),
        qkv_weight: vec![0.001; 3 * hidden * hidden],
        qkv_bias: Some(vec![0.0; 3 * hidden]),
        attn_output_weight: vec![0.001; hidden * hidden],
        attn_output_bias: Some(vec![0.0; hidden]),
        ffn_gate_weight: None, // GELU
        ffn_gate_bias: None,
        ffn_up_weight: vec![0.001; intermediate * hidden],
        ffn_up_bias: Some(vec![0.0; intermediate]),
        ffn_down_weight: vec![0.001; hidden * intermediate],
        ffn_down_bias: Some(vec![0.0; hidden]),
        ffn_norm_weight: Some(vec![1.0; hidden]),
        ffn_norm_bias: Some(vec![0.0; hidden]),
        attn_q_norm_weight: None,
        attn_k_norm_weight: None,
        linear_attn_z_weight: None,
        linear_attn_b_weight: None,
        linear_attn_a_weight: None,
        linear_attn_conv1d_weight: None,
        linear_attn_a_log: None,
        linear_attn_dt_bias: None,
        linear_attn_norm_weight: None,
        moe_gate_weight: None,
        moe_expert_gate_up: None,
        moe_expert_down: None,
        moe_shared_gate: None,
        moe_shared_up: None,
        moe_shared_down: None,
        moe_shared_expert_gate_weight: None,
    };

    let apr = AprTransformer {
        config: AprTransformerConfig {
            architecture: "phi2".to_string(),
            hidden_dim: hidden,
            num_layers: 1,
            num_heads: 4,
            num_kv_heads: 4,
            vocab_size: vocab,
            intermediate_dim: intermediate,
            context_length: 2048,
            rope_theta: 10000.0,
            eps: 1e-5,
            eos_token_id: None,
        ..Default::default()
        },
        token_embedding: vec![0.01; vocab * hidden],
        layers: vec![layer],
        output_norm_weight: vec![1.0; hidden],
        output_norm_bias: Some(vec![0.0; hidden]),
        lm_head_weight: vec![0.01; vocab * hidden],
        lm_head_bias: Some(vec![0.0; vocab]),
        lm_head_tied: false,
        q4k_layers: None,
        lm_head_weight_q6k: None,
        lm_head_weight_q4k: None,
    };

    let result = apr.forward(&[1, 2]);
    assert!(
        result.is_ok(),
        "GELU batch + trace: {}",
        result.unwrap_err()
    );
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::remove_var("REALIZE_TRACE");
    }
}

// ============================================================================
// forward (batch): partial Q4K + trace (batch forward F32 fallback trace)
// ============================================================================

#[test]
fn test_forward_batch_partial_q4k_with_trace() {
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::set_var("REALIZE_TRACE", "1");
    }

    let hidden = 32;
    let intermediate = 64;
    let vocab = 16;
    let heads = 4;
    let kv_heads = 4;
    let head_dim = hidden / heads;
    let kv_size = kv_heads * head_dim;

    let layer = AprTransformerLayer {
        attn_norm_weight: vec![1.0; hidden],
        attn_norm_bias: None,
        qkv_weight: vec![0.001; (hidden + 2 * kv_size) * hidden],
        qkv_bias: None,
        attn_output_weight: vec![0.001; hidden * hidden],
        attn_output_bias: None,
        ffn_gate_weight: Some(vec![0.001; intermediate * hidden]),
        ffn_gate_bias: None,
        ffn_up_weight: vec![0.001; intermediate * hidden],
        ffn_up_bias: None,
        ffn_down_weight: vec![0.001; hidden * intermediate],
        ffn_down_bias: None,
        ffn_norm_weight: Some(vec![1.0; hidden]),
        ffn_norm_bias: None,
        attn_q_norm_weight: None,
        attn_k_norm_weight: None,
        linear_attn_z_weight: None,
        linear_attn_b_weight: None,
        linear_attn_a_weight: None,
        linear_attn_conv1d_weight: None,
        linear_attn_a_log: None,
        linear_attn_dt_bias: None,
        linear_attn_norm_weight: None,
        moe_gate_weight: None,
        moe_expert_gate_up: None,
        moe_expert_down: None,
        moe_shared_gate: None,
        moe_shared_up: None,
        moe_shared_down: None,
        moe_shared_expert_gate_weight: None,
    };

    // Only attn_output Q4K populated — all others fall through to F32
    let q4k = Q4KLayerWeights {
        qkv_weight: None,
        attn_q_weight: None,
        attn_k_weight: None,
        attn_v_weight: None,
        attn_v_weight_q6k: None,
        attn_output_weight: Some(q4k_bytes(hidden, hidden)),
        ffn_gate_weight: None,
        ffn_up_weight: None,
        ffn_down_weight: None,
        ffn_down_weight_q6k: None,
        ffn_up_weight_q6k: None,
    };

    let apr = AprTransformer {
        config: AprTransformerConfig {
            architecture: "llama".to_string(),
            hidden_dim: hidden,
            num_layers: 1,
            num_heads: heads,
            num_kv_heads: kv_heads,
            vocab_size: vocab,
            intermediate_dim: intermediate,
            context_length: 2048,
            rope_theta: 10000.0,
            eps: 1e-5,
            eos_token_id: None,
        ..Default::default()
        },
        token_embedding: vec![0.01; vocab * hidden],
        layers: vec![layer],
        output_norm_weight: vec![1.0; hidden],
        output_norm_bias: None,
        lm_head_weight: vec![0.01; vocab * hidden],
        lm_head_bias: None,
        lm_head_tied: false,
        q4k_layers: Some(vec![q4k]),
        lm_head_weight_q6k: None,
        lm_head_weight_q4k: None,
    };

    let result = apr.forward(&[1, 2, 3]);
    assert!(
        result.is_ok(),
        "Partial Q4K batch + trace: {}",
        result.unwrap_err()
    );
    // SAFETY: Preconditions verified by caller or enclosing context
    unsafe {
        std::env::remove_var("REALIZE_TRACE");
    }
}