aprender-serve 0.33.0

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


// ============================================================================
// qkv_matmul_into() Tests
// ============================================================================

#[test]
fn test_qkv_matmul_into_fused() {
    let model = create_test_model(256, 100);
    let qkv_weight = create_q4k_test_data(256, 3 * 256);
    let qkv = OwnedQKVWeights::Fused(qkv_weight);
    let input = vec![1.0f32; 256];
    let mut output = vec![0.0f32; 3 * 256];

    let result = model.qkv_matmul_into(&input, &qkv, &mut output);
    assert!(result.is_ok());
}

#[test]
fn test_qkv_matmul_into_separate() {
    let model = create_test_model(256, 100);
    let q = create_q4k_test_data(256, 256);
    let k = create_q4k_test_data(256, 64);
    let v = create_q4k_test_data(256, 64);
    let qkv = OwnedQKVWeights::Separate { q, k, v };
    let input = vec![1.0f32; 256];
    let mut output = vec![0.0f32; 384];

    let result = model.qkv_matmul_into(&input, &qkv, &mut output);
    assert!(result.is_ok());
}

// ============================================================================
// layer_norm() Tests
// ============================================================================

#[test]
fn test_layer_norm_basic() {
    let model = create_test_model(64, 100);
    let input = vec![1.0f32; 64];
    let weight = vec![1.0f32; 64];

    let output = model.layer_norm(&input, &weight, None, 1e-5);
    assert_eq!(output.len(), 64);
}

#[test]
fn test_layer_norm_with_bias() {
    let model = create_test_model(64, 100);
    let input = vec![1.0f32; 64];
    let weight = vec![1.0f32; 64];
    let bias = vec![0.5f32; 64];

    let output = model.layer_norm(&input, &weight, Some(&bias), 1e-5);
    assert_eq!(output.len(), 64);
}

// ============================================================================
// add_bias() Tests
// ============================================================================

#[test]
fn test_add_bias() {
    let model = create_test_model(64, 100);
    let mut input = vec![1.0f32; 4];
    let bias = vec![0.5f32; 4];

    model.add_bias(&mut input, &bias);
    assert!(input.iter().all(|&x| (x - 1.5).abs() < f32::EPSILON));
}

#[test]
fn test_add_bias_zeros() {
    let model = create_test_model(64, 100);
    let mut input = vec![2.0f32; 4];
    let bias = vec![0.0f32; 4];

    model.add_bias(&mut input, &bias);
    assert!(input.iter().all(|&x| (x - 2.0).abs() < f32::EPSILON));
}

// ============================================================================
// gelu() Tests
// ============================================================================

#[test]
fn test_gelu_zeros() {
    let model = create_test_model(64, 100);
    let mut input = vec![0.0f32; 4];
    model.gelu(&mut input);
    // GELU(0) = 0
    assert!(input.iter().all(|&x| x.abs() < 1e-6));
}

#[test]
fn test_gelu_positive() {
    let model = create_test_model(64, 100);
    let mut input = vec![1.0f32; 4];
    model.gelu(&mut input);
    // GELU(1) ≈ 0.841
    assert!(input.iter().all(|&x| (x - 0.841).abs() < 0.01));
}

#[test]
fn test_gelu_negative() {
    let model = create_test_model(64, 100);
    let mut input = vec![-1.0f32; 4];
    model.gelu(&mut input);
    // GELU(-1) ≈ -0.159
    assert!(input.iter().all(|&x| (x - (-0.159)).abs() < 0.01));
}

// ============================================================================
// fused_rmsnorm_qkv_matmul() Tests
// ============================================================================

#[test]
fn test_fused_rmsnorm_qkv_matmul_fused() {
    let model = create_test_model(256, 100);
    let qkv_weight = create_q4k_test_data(256, 3 * 256);
    let qkv = OwnedQKVWeights::Fused(qkv_weight);
    let input = vec![1.0f32; 256];
    let norm_weight = vec![1.0f32; 256];

    let result = model.fused_rmsnorm_qkv_matmul(&input, &norm_weight, 1e-5, &qkv);
    assert!(result.is_ok());
    let output = result.unwrap();
    assert_eq!(output.len(), 3 * 256);
}

#[test]
fn test_fused_rmsnorm_qkv_matmul_separate() {
    let model = create_test_model(256, 100);
    let q = create_q4k_test_data(256, 256);
    let k = create_q4k_test_data(256, 64);
    let v = create_q4k_test_data(256, 64);
    let qkv = OwnedQKVWeights::Separate { q, k, v };
    let input = vec![1.0f32; 256];
    let norm_weight = vec![1.0f32; 256];

    let result = model.fused_rmsnorm_qkv_matmul(&input, &norm_weight, 1e-5, &qkv);
    assert!(result.is_ok());
    let output = result.unwrap();
    assert_eq!(output.len(), 384);
}

// ============================================================================
// fused_rmsnorm_lm_head() Tests
// ============================================================================

#[test]
fn test_fused_rmsnorm_lm_head_q4k() {
    let model = create_test_model(256, 100);
    let input = vec![1.0f32; 256];

    let result = model.fused_rmsnorm_lm_head(&input);
    assert!(result.is_ok());
    let output = result.unwrap();
    assert_eq!(output.len(), 100); // vocab_size
}

// ============================================================================
// fused_rmsnorm_ffn_up_gate() Tests
// ============================================================================

#[test]
fn test_fused_rmsnorm_ffn_up_gate_q4_0() {
    let model = create_test_model(64, 100);
    let up_weight = create_q4_0_test_data(64, 256);
    let gate_weight = create_q4_0_test_data(64, 256);
    let input = vec![1.0f32; 64];
    let norm_weight = vec![1.0f32; 64];

    let result =
        model.fused_rmsnorm_ffn_up_gate(&input, &norm_weight, 1e-5, &up_weight, &gate_weight);
    assert!(result.is_ok());
    let (up_out, gate_out) = result.unwrap();
    assert_eq!(up_out.len(), 256);
    assert_eq!(gate_out.len(), 256);
}

#[test]
fn test_fused_rmsnorm_ffn_up_gate_q4k_fallback() {
    let model = create_test_model(256, 100);
    let up_weight = create_q4k_test_data(256, 1024);
    let gate_weight = create_q4k_test_data(256, 1024);
    let input = vec![1.0f32; 256];
    let norm_weight = vec![1.0f32; 256];

    let result =
        model.fused_rmsnorm_ffn_up_gate(&input, &norm_weight, 1e-5, &up_weight, &gate_weight);
    assert!(result.is_ok());
    let (up_out, gate_out) = result.unwrap();
    assert_eq!(up_out.len(), 1024);
    assert_eq!(gate_out.len(), 1024);
}

// ============================================================================
// qkv_matmul_q8k_into() Tests
// ============================================================================

#[test]
fn test_qkv_matmul_q8k_into_fused() {
    let model = create_test_model(256, 100);
    let qkv_weight = create_q4k_test_data(256, 3 * 256);
    let qkv = OwnedQKVWeights::Fused(qkv_weight);
    let input = vec![1.0f32; 256];
    let mut output = vec![0.0f32; 3 * 256];
    let scales = vec![1.0f32; 8];
    let quants = vec![0i8; 256];

    // Currently falls back to regular qkv_matmul_into
    let result = model.qkv_matmul_q8k_into(&input, &qkv, &mut output, &scales, &quants);
    assert!(result.is_ok());
}

// ============================================================================
// Edge Case Tests
// ============================================================================

// ============================================================================
// GH-242: F32 / F16 fused_matmul tests (type 0 / type 1)
// ============================================================================

/// Create F32 weight tensor — raw little-endian f32 bytes
fn create_f32_test_data(in_dim: usize, out_dim: usize) -> OwnedQuantizedTensor {
    let mut data = Vec::with_capacity(out_dim * in_dim * 4);
    for row in 0..out_dim {
        for col in 0..in_dim {
            // Identity-like: 0.1 on diagonal, 0.0 elsewhere
            let val = if row == col { 0.1f32 } else { 0.0f32 };
            data.extend_from_slice(&val.to_le_bytes());
        }
    }
    OwnedQuantizedTensor {
        data,
        in_dim,
        out_dim,
        qtype: GGUF_TYPE_F32,
    }
}

/// Create F16 weight tensor — raw little-endian f16 bytes
fn create_f16_test_data(in_dim: usize, out_dim: usize) -> OwnedQuantizedTensor {
    let mut data = Vec::with_capacity(out_dim * in_dim * 2);
    for row in 0..out_dim {
        for col in 0..in_dim {
            let val = if row == col { 0.1f32 } else { 0.0f32 };
            let bits = half::f16::from_f32(val).to_bits();
            data.extend_from_slice(&bits.to_le_bytes());
        }
    }
    OwnedQuantizedTensor {
        data,
        in_dim,
        out_dim,
        qtype: GGUF_TYPE_F16,
    }
}

/// Create a model with F32 weights (GH-242 regression test)
fn create_f32_test_model(hidden_dim: usize, vocab_size: usize) -> OwnedQuantizedModel {
    let config = test_config(hidden_dim, vocab_size);
    let intermediate_dim = config.intermediate_dim;

    let qkv_weight = create_f32_test_data(hidden_dim, 3 * hidden_dim);
    let attn_output_weight = create_f32_test_data(hidden_dim, hidden_dim);
    let ffn_up_weight = create_f32_test_data(hidden_dim, intermediate_dim);
    let ffn_down_weight = create_f32_test_data(intermediate_dim, hidden_dim);

    let layer = crate::gguf::OwnedQuantizedLayer {
        attn_norm_weight: vec![1.0f32; hidden_dim],
        attn_norm_bias: None,
        qkv_weight: OwnedQKVWeights::Fused(qkv_weight),
        qkv_bias: None,
        attn_output_weight,
        attn_output_bias: None,
        ffn_up_weight,
        ffn_up_bias: None,
        ffn_down_weight,
        ffn_down_bias: None,
        ffn_gate_weight: None,
        ffn_gate_bias: None,
        ffn_norm_weight: Some(vec![1.0f32; hidden_dim]),
        ffn_norm_bias: None,
        attn_q_norm_weight: None,
        attn_k_norm_weight: None,
    };

    let lm_head_weight = create_f32_test_data(hidden_dim, vocab_size);

    OwnedQuantizedModel {
        config,
        token_embedding: vec![0.1f32; vocab_size * hidden_dim],
        position_embedding: None,
        layers: vec![layer],
        encoder_layers: vec![],
        encoder_output_norm_weight: None,
        encoder_output_norm_bias: None,
        output_norm_weight: vec![1.0f32; hidden_dim],
        output_norm_bias: None,
        lm_head_weight,
        lm_head_bias: None,
        #[cfg(feature = "cuda")]
        cuda_executor: None,
        #[cfg(feature = "cuda")]
        cuda_kernel_count: std::sync::atomic::AtomicU64::new(0),
        #[cfg(feature = "cuda")]
        cached_weight_names: std::sync::Mutex::new(std::collections::HashSet::new()),
    }
}

#[test]
fn test_fused_matmul_f32_single_token() {
    // GH-242: F32 weights (type 0) must not error in fused_matmul
    let model = create_f32_test_model(64, 100);
    let input = vec![1.0f32; 64];
    let weight = create_f32_test_data(64, 32);
    let result = model.fused_matmul(&input, &weight);
    assert!(result.is_ok(), "F32 fused_matmul failed: {:?}", result.err());
    assert_eq!(result.unwrap().len(), 32);
}

#[test]
fn test_fused_matmul_f32_multi_token() {
    // GH-242: F32 with seq_len > 1
    let model = create_f32_test_model(64, 100);
    let input = vec![1.0f32; 64 * 3]; // 3 tokens
    let weight = create_f32_test_data(64, 32);
    let result = model.fused_matmul(&input, &weight);
    assert!(result.is_ok(), "F32 multi-token fused_matmul failed: {:?}", result.err());
    assert_eq!(result.unwrap().len(), 32 * 3);
}

#[test]
fn test_fused_matmul_f16_single_token() {
    // GH-242: F16 weights (type 1) must not error in fused_matmul
    let model = create_f32_test_model(64, 100);
    let input = vec![1.0f32; 64];
    let weight = create_f16_test_data(64, 32);
    let result = model.fused_matmul(&input, &weight);
    assert!(result.is_ok(), "F16 fused_matmul failed: {:?}", result.err());
    assert_eq!(result.unwrap().len(), 32);
}

#[test]
fn test_fused_matmul_f32_correctness() {
    // GH-242: Verify F32 produces correct output (identity-like weight)
    let model = create_f32_test_model(8, 16);
    let mut input = vec![0.0f32; 8];
    input[0] = 5.0;
    input[1] = 3.0;
    let weight = create_f32_test_data(8, 8); // 0.1 on diagonal
    let result = model.fused_matmul(&input, &weight).unwrap();
    // output[0] = input[0] * 0.1 = 0.5, output[1] = input[1] * 0.1 = 0.3
    assert!((result[0] - 0.5).abs() < 1e-5, "Expected 0.5, got {}", result[0]);
    assert!((result[1] - 0.3).abs() < 1e-5, "Expected 0.3, got {}", result[1]);
}

#[test]
fn test_fused_matmul_into_f32() {
    // GH-242: fused_matmul_into delegates F32 to fused_matmul
    let model = create_f32_test_model(64, 100);
    let input = vec![1.0f32; 64];
    let weight = create_f32_test_data(64, 32);
    let mut output = vec![0.0f32; 32];
    let result = model.fused_matmul_into(&input, &weight, &mut output);
    assert!(result.is_ok(), "F32 fused_matmul_into failed: {:?}", result.err());
}

#[test]
fn test_embed_max_token() {
    let model = create_test_model(64, 1000);
    let embeddings = model.embed(&[999]);
    assert_eq!(embeddings.len(), 64);
}

#[test]
fn test_embed_into_larger_buffer() {
    let model = create_test_model(64, 100);
    let mut output = vec![9.9f32; 128]; // Buffer larger than needed
    model.embed_into(0, &mut output);
    // First hidden_dim elements should be filled
    assert!(output[..64].iter().all(|&x| (x - 0.1).abs() < f32::EPSILON));
}