irithyll 10.0.0

Streaming ML in Rust -- gradient boosted trees, neural architectures (TTT/KAN/MoE/Mamba/SNN), AutoML, kernel methods, and composable pipelines
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

//! Benchmarks for int16 quantized inference vs f32 packed inference.
//!
//! Measures latency, throughput, accuracy degradation, and binary size for the
//! quantized export path.

use criterion::{black_box, criterion_group, criterion_main, BenchmarkId, Criterion, Throughput};
use irithyll::ensemble::config::SGBTConfig;
use irithyll::ensemble::SGBT;
use irithyll::export_embedded::{export_packed, export_packed_i16};
use irithyll::sample::Sample;
use irithyll_core::{EnsembleView, QuantizedEnsembleView};

// ---------------------------------------------------------------------------
// Deterministic helpers (no rand dependency)
// ---------------------------------------------------------------------------

fn xorshift64(state: &mut u64) -> f64 {
    *state ^= *state << 13;
    *state ^= *state >> 7;
    *state ^= *state << 17;
    (*state as f64) / (u64::MAX as f64)
}

fn train_bench_model(
    n_steps: usize,
    n_features: usize,
    max_depth: usize,
    n_samples: usize,
) -> SGBT {
    let config = SGBTConfig::builder()
        .n_steps(n_steps)
        .learning_rate(0.01)
        .grace_period(20)
        .max_depth(max_depth)
        .n_bins(32)
        .build()
        .unwrap();
    let mut model = SGBT::new(config);
    let mut rng: u64 = 0xBEEF_CAFE;
    for _ in 0..n_samples {
        let features: Vec<f64> = (0..n_features)
            .map(|_| xorshift64(&mut rng) * 10.0 - 5.0)
            .collect();
        let target = features
            .iter()
            .enumerate()
            .fold(0.0, |acc, (i, &f)| acc + (i as f64 + 1.0) * f);
        model.train_one(&Sample::new(features, target));
    }
    model
}

fn generate_f32_samples(n: usize, n_features: usize) -> Vec<Vec<f32>> {
    let mut rng: u64 = 0xDEAD_1337;
    (0..n)
        .map(|_| {
            (0..n_features)
                .map(|_| {
                    let v = xorshift64(&mut rng) * 10.0 - 5.0;
                    v as f32
                })
                .collect()
        })
        .collect()
}

// ---------------------------------------------------------------------------
// Group 1: quantized_predict_vs_f32
//
// Same model exported to both f32 packed and int16 quantized formats.
// Sweep n_trees in [10, 50, 100, 500].
// Hold max_depth=4, n_features=10.
// ---------------------------------------------------------------------------

fn quantized_predict_vs_f32(c: &mut Criterion) {
    let mut group = c.benchmark_group("quantized_vs_f32");
    let n_features = 10;
    let max_depth = 4;
    let n_samples = 1000;

    for n_trees in [10, 50, 100, 500] {
        let model = train_bench_model(n_trees, n_features, max_depth, n_samples);

        let packed_f32 = export_packed(&model, n_features);
        let view_f32 = EnsembleView::from_bytes(&packed_f32).unwrap();

        let packed_i16 = export_packed_i16(&model, n_features);
        let view_i16 = QuantizedEnsembleView::from_bytes(&packed_i16).unwrap();

        let features: Vec<f32> = (0..n_features).map(|i| (i as f32) * 0.1).collect();

        group.bench_with_input(BenchmarkId::new("f32_packed", n_trees), &n_trees, |b, _| {
            b.iter(|| view_f32.predict(black_box(&features)))
        });
        group.bench_with_input(
            BenchmarkId::new("i16_quantized", n_trees),
            &n_trees,
            |b, _| b.iter(|| view_i16.predict(black_box(&features))),
        );
    }
    group.finish();
}

// ---------------------------------------------------------------------------
// Group 2: quantized_batch
//
// Batch prediction comparison f32 vs i16.
// Batch sizes: [1_000, 10_000, 100_000].
// Hold n_trees=50, max_depth=4, n_features=10.
// ---------------------------------------------------------------------------

fn quantized_batch(c: &mut Criterion) {
    let mut group = c.benchmark_group("quantized_batch");
    let n_trees = 50;
    let n_features = 10;
    let max_depth = 4;
    let n_samples = 1000;

    let model = train_bench_model(n_trees, n_features, max_depth, n_samples);

    let packed_f32 = export_packed(&model, n_features);
    let view_f32 = EnsembleView::from_bytes(&packed_f32).unwrap();

    let packed_i16 = export_packed_i16(&model, n_features);
    let view_i16 = QuantizedEnsembleView::from_bytes(&packed_i16).unwrap();

    for batch_size in [1_000u64, 10_000, 100_000] {
        let samples = generate_f32_samples(batch_size as usize, n_features);
        let sample_refs: Vec<&[f32]> = samples.iter().map(|s| s.as_slice()).collect();
        let mut out_f32 = vec![0.0f32; batch_size as usize];
        let mut out_i16 = vec![0.0f32; batch_size as usize];

        group.throughput(Throughput::Elements(batch_size));

        group.bench_with_input(
            BenchmarkId::new("f32_packed", batch_size),
            &batch_size,
            |b, _| {
                b.iter(|| {
                    view_f32.predict_batch(black_box(&sample_refs), &mut out_f32);
                    black_box(&out_f32);
                });
            },
        );

        group.bench_with_input(
            BenchmarkId::new("i16_quantized", batch_size),
            &batch_size,
            |b, _| {
                b.iter(|| {
                    view_i16.predict_batch(black_box(&sample_refs), &mut out_i16);
                    black_box(&out_i16);
                });
            },
        );
    }
    group.finish();
}

// ---------------------------------------------------------------------------
// Group 3: quantized_accuracy
//
// Measure max |delta| between f32 and i16 predictions across 10K samples.
// Sweep n_trees in [10, 50, 100, 500].
// Not primarily a latency bench -- a correctness validation bench.
// ---------------------------------------------------------------------------

fn quantized_accuracy(c: &mut Criterion) {
    let mut group = c.benchmark_group("quantized_accuracy");
    let n_features = 10;
    let max_depth = 4;
    let n_train_samples = 1000;
    let n_test_samples = 10_000;

    for n_trees in [10, 50, 100, 500] {
        let model = train_bench_model(n_trees, n_features, max_depth, n_train_samples);

        let packed_f32 = export_packed(&model, n_features);
        let view_f32 = EnsembleView::from_bytes(&packed_f32).unwrap();

        let packed_i16 = export_packed_i16(&model, n_features);
        let view_i16 = QuantizedEnsembleView::from_bytes(&packed_i16).unwrap();

        group.bench_with_input(BenchmarkId::from_parameter(n_trees), &n_trees, |b, _| {
            b.iter(|| {
                let mut max_delta: f32 = 0.0;
                let mut rng: u64 = 0xACE0_FACE;
                for _ in 0..n_test_samples {
                    let features: Vec<f32> = (0..n_features)
                        .map(|_| {
                            let v = xorshift64(&mut rng);
                            v as f32 * 10.0 - 5.0
                        })
                        .collect();
                    let f32_pred = view_f32.predict(&features);
                    let i16_pred = view_i16.predict(&features);
                    let delta = (f32_pred - i16_pred).abs();
                    if delta > max_delta {
                        max_delta = delta;
                    }
                }
                black_box(max_delta)
            });
        });

        // Print the accuracy result for each tree count (visible in criterion output)
        let mut max_delta: f32 = 0.0;
        let mut sum_delta: f64 = 0.0;
        let mut rng: u64 = 0xACE0_FACE;
        for _ in 0..n_test_samples {
            let features: Vec<f32> = (0..n_features)
                .map(|_| {
                    let v = xorshift64(&mut rng);
                    v as f32 * 10.0 - 5.0
                })
                .collect();
            let f32_pred = view_f32.predict(&features);
            let i16_pred = view_i16.predict(&features);
            let delta = (f32_pred - i16_pred).abs();
            sum_delta += delta as f64;
            if delta > max_delta {
                max_delta = delta;
            }
        }
        let mean_delta = sum_delta / n_test_samples as f64;
        eprintln!(
            "[ACCURACY] {}t: max_delta={:.6}, mean_delta={:.6}",
            n_trees, max_delta, mean_delta
        );
    }
    group.finish();
}

// ---------------------------------------------------------------------------
// Group 4: quantized_model_size
//
// Export speed + binary size comparison for both formats.
// Sweep n_trees in [10, 50, 100, 500].
// Hold max_depth=4, n_features=10.
// ---------------------------------------------------------------------------

fn quantized_model_size(c: &mut Criterion) {
    let mut group = c.benchmark_group("quantized_model_size");
    let n_features = 10;
    let max_depth = 4;
    let n_samples = 1000;

    for n_trees in [10, 50, 100, 500] {
        let model = train_bench_model(n_trees, n_features, max_depth, n_samples);

        let packed_f32 = export_packed(&model, n_features);
        let packed_i16 = export_packed_i16(&model, n_features);

        let f32_nodes = EnsembleView::from_bytes(&packed_f32).unwrap().total_nodes();
        let i16_nodes = QuantizedEnsembleView::from_bytes(&packed_i16)
            .unwrap()
            .total_nodes();

        eprintln!(
            "[SIZE] {}t: f32={} bytes ({} nodes), i16={} bytes ({} nodes), ratio={:.2}x",
            n_trees,
            packed_f32.len(),
            f32_nodes,
            packed_i16.len(),
            i16_nodes,
            packed_f32.len() as f64 / packed_i16.len() as f64
        );

        // Bench export speed for f32
        group.bench_with_input(BenchmarkId::new("export_f32", n_trees), &n_trees, |b, _| {
            b.iter(|| export_packed(black_box(&model), n_features))
        });

        // Bench export speed for i16
        group.bench_with_input(BenchmarkId::new("export_i16", n_trees), &n_trees, |b, _| {
            b.iter(|| export_packed_i16(black_box(&model), n_features))
        });
    }
    group.finish();
}

// ---------------------------------------------------------------------------
// Criterion wiring
// ---------------------------------------------------------------------------

criterion_group!(
    quantized_benches,
    quantized_predict_vs_f32,
    quantized_batch,
    quantized_accuracy,
    quantized_model_size,
);
criterion_main!(quantized_benches);