aprender-serve 0.64.0

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


#[test]
fn test_fused_q4k_tiled_matvec_large() {
    // RED: Test with larger dimensions to exercise tiling
    use crate::quantize::fused_q4k_tiled_matvec;

    // 128 output dimensions, 512 input dimensions (2 super-blocks per row)
    let in_dim = 512;
    let out_dim = 128;
    let bytes_per_row = 2 * 144; // 2 super-blocks × 144 bytes

    // Create weight data
    let mut weight_data = Vec::with_capacity(out_dim * bytes_per_row);
    for row in 0..out_dim {
        for sb in 0..2 {
            let d = 1.0 + (row as f32) * 0.01 + (sb as f32) * 0.001;
            weight_data.extend_from_slice(&half::f16::from_f32(d).to_bits().to_le_bytes());
            weight_data.extend_from_slice(&half::f16::from_f32(0.0).to_bits().to_le_bytes());
            for i in 0..12 {
                weight_data.push(((row * 3 + sb * 5 + i) % 64) as u8);
            }
            for i in 0..128 {
                weight_data.push(((row * 7 + sb * 11 + i) % 256) as u8);
            }
        }
    }

    // Activations
    let activations: Vec<f32> = (0..in_dim).map(|i| (i as f32 * 0.005).cos()).collect();

    // Reference
    let mut reference = Vec::with_capacity(out_dim);
    for row in 0..out_dim {
        let row_start = row * bytes_per_row;
        let row_data = &weight_data[row_start..row_start + bytes_per_row];
        let dot = fused_q4k_dot_simd(row_data, &activations).expect("test");
        reference.push(dot);
    }

    // Tiled with default tile size (64)
    let tiled =
        fused_q4k_tiled_matvec(&weight_data, &activations, in_dim, out_dim, None).expect("test");

    assert_eq!(tiled.len(), out_dim);
    for i in 0..out_dim {
        assert_ulp_eq(
            tiled[i],
            reference[i],
            8,
            &format!("tiled_matvec_large output {}", i),
        );
    }
}

#[test]
fn test_fused_q4k_tiled_matvec_custom_tile_size() {
    // RED: Test that different tile sizes produce same results
    use crate::quantize::fused_q4k_tiled_matvec;

    let in_dim = 256;
    let out_dim = 100;

    // Create weight data
    let mut weight_data = Vec::with_capacity(out_dim * 144);
    for row in 0..out_dim {
        let d = 1.0 + (row as f32) * 0.02;
        weight_data.extend_from_slice(&half::f16::from_f32(d).to_bits().to_le_bytes());
        weight_data.extend_from_slice(&half::f16::from_f32(0.1).to_bits().to_le_bytes());
        for i in 0..12 {
            weight_data.push(((row + i) % 64) as u8);
        }
        for i in 0..128 {
            weight_data.push(((row * 2 + i) % 256) as u8);
        }
    }

    let activations: Vec<f32> = (0..in_dim).map(|i| i as f32 * 0.01).collect();

    // Test with different tile sizes
    let tile_sizes = [1, 8, 16, 32, 64, 100, 128];
    let reference =
        fused_q4k_tiled_matvec(&weight_data, &activations, in_dim, out_dim, Some(1)).expect("test");

    for &tile_size in &tile_sizes[1..] {
        let result =
            fused_q4k_tiled_matvec(&weight_data, &activations, in_dim, out_dim, Some(tile_size))
                .expect("test");
        assert_eq!(result.len(), out_dim);
        for i in 0..out_dim {
            assert_ulp_eq(
                result[i],
                reference[i],
                4,
                &format!("tile_size={} output {}", tile_size, i),
            );
        }
    }
}

#[test]
fn test_fused_q4k_tiled_matvec_error_handling() {
    // RED: Test error cases
    use crate::quantize::fused_q4k_tiled_matvec;

    // Weight data too small
    let small_data = vec![0u8; 100];
    let activations = vec![0.0f32; 256];
    assert!(fused_q4k_tiled_matvec(&small_data, &activations, 256, 4, None).is_err());

    // Activation length mismatch
    let weight_data = vec![0u8; 4 * 144];
    let bad_activations = vec![0.0f32; 128];
    assert!(fused_q4k_tiled_matvec(&weight_data, &bad_activations, 256, 4, None).is_err());
}

// -------------------------------------------------------------------------
// Phase 2: Parallel Matrix-Vector Multiplication Tests
// -------------------------------------------------------------------------

#[test]
fn test_fused_q4k_parallel_matvec_basic() {
    // RED: Test parallel matvec produces same results as sequential
    use crate::quantize::fused_q4k_parallel_matvec;

    let in_dim = 256;
    let out_dim = 64;

    // Create weight data
    let mut weight_data = Vec::with_capacity(out_dim * 144);
    for row in 0..out_dim {
        let d = 0.5 + (row as f32) * 0.01;
        weight_data.extend_from_slice(&half::f16::from_f32(d).to_bits().to_le_bytes());
        weight_data.extend_from_slice(&half::f16::from_f32(0.05).to_bits().to_le_bytes());
        for i in 0..12 {
            weight_data.push(((row * 7 + i) % 64) as u8);
        }
        for i in 0..128 {
            weight_data.push(((row * 13 + i) % 256) as u8);
        }
    }

    let activations: Vec<f32> = (0..in_dim).map(|i| (i as f32 * 0.01).sin()).collect();

    // Reference: sequential computation
    let mut reference = Vec::with_capacity(out_dim);
    for row in 0..out_dim {
        let row_start = row * 144;
        let row_data = &weight_data[row_start..row_start + 144];
        let dot = fused_q4k_dot_simd(row_data, &activations).expect("test");
        reference.push(dot);
    }

    // Parallel result (uses Q8K activation quantization internally — Refs realizar#96)
    let parallel =
        fused_q4k_parallel_matvec(&weight_data, &activations, in_dim, out_dim).expect("test");

    assert_eq!(parallel.len(), out_dim);
    // Contract: cpu-q4k-activation-quant-v1.yaml — Q8K error < 0.1% relative to f32
    for i in 0..out_dim {
        let rel_err = if reference[i].abs() > 1e-6 {
            (parallel[i] - reference[i]).abs() / reference[i].abs()
        } else {
            (parallel[i] - reference[i]).abs()
        };
        assert!(
            rel_err < 0.002,
            "parallel_matvec output {}: actual={}, expected={}, rel_err={:.4}% > 0.2%",
            i, parallel[i], reference[i], rel_err * 100.0
        );
    }
}

#[test]
fn test_fused_q4k_parallel_matvec_large() {
    // RED: Test with larger dimensions typical of real models
    use crate::quantize::fused_q4k_parallel_matvec;

    let in_dim = 512;
    let out_dim = 256;
    let bytes_per_row = 2 * 144; // 2 super-blocks × 144 bytes

    // Create weight data
    let mut weight_data = Vec::with_capacity(out_dim * bytes_per_row);
    for row in 0..out_dim {
        for sb in 0..2 {
            let d = 1.0 + (row as f32) * 0.005 + (sb as f32) * 0.001;
            weight_data.extend_from_slice(&half::f16::from_f32(d).to_bits().to_le_bytes());
            weight_data.extend_from_slice(&half::f16::from_f32(0.0).to_bits().to_le_bytes());
            for i in 0..12 {
                weight_data.push(((row * 3 + sb * 5 + i) % 64) as u8);
            }
            for i in 0..128 {
                weight_data.push(((row * 7 + sb * 11 + i) % 256) as u8);
            }
        }
    }

    let activations: Vec<f32> = (0..in_dim).map(|i| (i as f32 * 0.003).cos()).collect();

    // Reference (f32 dot product path)
    let mut reference = Vec::with_capacity(out_dim);
    for row in 0..out_dim {
        let row_start = row * bytes_per_row;
        let row_data = &weight_data[row_start..row_start + bytes_per_row];
        let dot = fused_q4k_dot_simd(row_data, &activations).expect("test");
        reference.push(dot);
    }

    // Parallel result (uses Q8K activation quantization internally — Refs realizar#96)
    let parallel =
        fused_q4k_parallel_matvec(&weight_data, &activations, in_dim, out_dim).expect("test");

    assert_eq!(parallel.len(), out_dim);
    // Contract: cpu-q4k-activation-quant-v1.yaml — Q8K error < 0.1% relative to f32
    for i in 0..out_dim {
        let rel_err = if reference[i].abs() > 1e-6 {
            (parallel[i] - reference[i]).abs() / reference[i].abs()
        } else {
            (parallel[i] - reference[i]).abs()
        };
        assert!(
            rel_err < 0.002,
            "parallel_matvec_large output {}: actual={}, expected={}, rel_err={:.4}% > 0.2%",
            i, parallel[i], reference[i], rel_err * 100.0
        );
    }
}

#[test]
fn test_fused_q5k_parallel_matvec_basic() {
    // RED: Test Q5_K parallel matvec
    use crate::quantize::fused_q5k_parallel_matvec;

    let in_dim = 256;
    let out_dim = 32;
    let bytes_per_row = 176;

    // Create weight data
    let mut weight_data = Vec::with_capacity(out_dim * bytes_per_row);
    for row in 0..out_dim {
        let d = 0.5 + (row as f32) * 0.02;
        weight_data.extend_from_slice(&half::f16::from_f32(d).to_bits().to_le_bytes());
        weight_data.extend_from_slice(&half::f16::from_f32(0.05).to_bits().to_le_bytes());
        // scales (12 bytes)
        for i in 0..12 {
            weight_data.push(((row * 5 + i) % 64) as u8);
        }
        // qh (32 bytes)
        for i in 0..32 {
            weight_data.push(((row * 3 + i) % 256) as u8);
        }
        // qs (128 bytes)
        for i in 0..128 {
            weight_data.push(((row * 11 + i) % 256) as u8);
        }
    }

    let activations: Vec<f32> = (0..in_dim).map(|i| (i as f32 * 0.01).sin()).collect();

    // Reference
    let mut reference = Vec::with_capacity(out_dim);
    for row in 0..out_dim {
        let row_start = row * bytes_per_row;
        let row_data = &weight_data[row_start..row_start + bytes_per_row];
        let dot = fused_q5k_dot_simd(row_data, &activations).expect("test");
        reference.push(dot);
    }

    // Parallel result
    let parallel =
        fused_q5k_parallel_matvec(&weight_data, &activations, in_dim, out_dim).expect("test");

    assert_eq!(parallel.len(), out_dim);
    for i in 0..out_dim {
        assert_ulp_eq(
            parallel[i],
            reference[i],
            4,
            &format!("q5k_parallel output {}", i),
        );
    }
}

#[test]
fn test_fused_q6k_parallel_matvec_basic() {
    // RED: Test Q6_K parallel matvec
    use crate::quantize::fused_q6k_parallel_matvec;

    let in_dim = 256;
    let out_dim = 32;
    let bytes_per_row = 210;

    // Create weight data (Q6_K layout: ql + qh + scales + d)
    let mut weight_data = Vec::with_capacity(out_dim * bytes_per_row);
    for row in 0..out_dim {
        // ql: 128 bytes
        for i in 0..128 {
            weight_data.push(((row * 7 + i) % 256) as u8);
        }
        // qh: 64 bytes
        for i in 0..64 {
            weight_data.push(((row * 3 + i) % 256) as u8);
        }
        // scales: 16 bytes (i8)
        for i in 0..16 {
            weight_data.push(((row + i) % 128) as u8);
        }
        // d: 2 bytes (f16)
        let d = 0.5 + (row as f32) * 0.02;
        weight_data.extend_from_slice(&half::f16::from_f32(d).to_bits().to_le_bytes());
    }

    let activations: Vec<f32> = (0..in_dim).map(|i| (i as f32 * 0.01).sin()).collect();

    // Reference
    let mut reference = Vec::with_capacity(out_dim);
    for row in 0..out_dim {
        let row_start = row * bytes_per_row;
        let row_data = &weight_data[row_start..row_start + bytes_per_row];
        let dot = fused_q6k_dot_simd(row_data, &activations).expect("test");
        reference.push(dot);
    }

    // Parallel result
    let parallel =
        fused_q6k_parallel_matvec(&weight_data, &activations, in_dim, out_dim).expect("test");

    assert_eq!(parallel.len(), out_dim);
    for i in 0..out_dim {
        assert_ulp_eq(
            parallel[i],
            reference[i],
            4,
            &format!("q6k_parallel output {}", i),
        );
    }
}

#[test]
fn test_fused_parallel_matvec_error_handling() {
    // RED: Test error cases for parallel matvec
    use crate::quantize::fused_q4k_parallel_matvec;

    // Weight data too small
    let small_data = vec![0u8; 100];
    let activations = vec![0.0f32; 256];
    assert!(fused_q4k_parallel_matvec(&small_data, &activations, 256, 4).is_err());

    // Activation length mismatch
    let weight_data = vec![0u8; 4 * 144];
    let bad_activations = vec![0.0f32; 128];
    assert!(fused_q4k_parallel_matvec(&weight_data, &bad_activations, 256, 4).is_err());
}