oxicuda-quant 0.1.2

GPU-accelerated quantization and model compression engine for OxiCUDA
Documentation 
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

use std::hint::black_box;

use criterion::{BenchmarkId, Criterion, criterion_group, criterion_main};
use oxicuda_quant::pruning::magnitude::{MagnitudeNorm, MagnitudePruner};
use oxicuda_quant::scheme::minmax::{MinMaxQuantizer, QuantGranularity, QuantScheme};
use oxicuda_quant::scheme::nf4::Nf4Quantizer;

// --- data generators ---------------------------------------------------------

fn make_weights(n: usize) -> Vec<f32> {
    // Pseudo-random normally-distributed-ish weights using simple LCG.
    let mut x: u64 = 0xdeadbeef_cafebabe;
    (0..n)
        .map(|_| {
            x = x.wrapping_mul(6_364_136_223_846_793_005).wrapping_add(1);
            // Box-Muller-ish: combine two uniform draws
            let u1 = (x >> 40) as f32 / (1u32 << 24) as f32;
            let u2 = ((x >> 16) & 0xFF_FFFF) as f32 / (1u32 << 24) as f32;
            (u1 - 0.5) * 0.2 + (u2 - 0.5) * 0.02
        })
        .collect()
}

// --- MinMax INT8 symmetric ---------------------------------------------------

fn bench_int8_symmetric(c: &mut Criterion) {
    let mut group = c.benchmark_group("quant_int8_symmetric");
    let q = MinMaxQuantizer::int8_symmetric();

    let sizes: &[(&str, usize)] = &[
        ("1k", 1_024),
        ("16k", 16_384),
        ("256k", 262_144),
        ("1m", 1_048_576),
        ("4m", 4_194_304),
    ];

    for &(name, n) in sizes {
        let data = make_weights(n);
        let params = q.calibrate(&data).unwrap();

        group.bench_with_input(BenchmarkId::new("calibrate", name), &n, |b, _| {
            b.iter(|| black_box(q.calibrate(&data).unwrap()));
        });

        group.bench_with_input(BenchmarkId::new("quantize", name), &n, |b, _| {
            b.iter(|| black_box(q.quantize(&data, &params).unwrap()));
        });

        group.bench_with_input(BenchmarkId::new("dequantize", name), &n, |b, _| {
            let codes = q.quantize(&data, &params).unwrap();
            b.iter(|| black_box(q.dequantize(&codes, &params)));
        });

        group.bench_with_input(BenchmarkId::new("round_trip", name), &n, |b, _| {
            b.iter(|| {
                let p = q.calibrate(&data).unwrap();
                let codes = q.quantize(&data, &p).unwrap();
                black_box(q.dequantize(&codes, &p))
            });
        });
    }
    group.finish();
}

// --- MinMax INT4 per-group ---------------------------------------------------

fn bench_int4_per_group(c: &mut Criterion) {
    let mut group = c.benchmark_group("quant_int4_per_group");

    let sizes: &[(&str, usize, usize)] = &[
        ("1k_g128", 1_024, 128),
        ("16k_g128", 16_384, 128),
        ("64k_g128", 65_536, 128),
        ("256k_g64", 262_144, 64),
    ];

    for &(name, n, gs) in sizes {
        let data = make_weights(n);
        let q = MinMaxQuantizer::int4_per_group(gs);

        group.bench_with_input(BenchmarkId::new("calibrate", name), &n, |b, _| {
            b.iter(|| black_box(q.calibrate(&data).unwrap()));
        });

        group.bench_with_input(BenchmarkId::new("quantize_grouped", name), &n, |b, _| {
            let params = q.calibrate(&data).unwrap();
            b.iter(|| black_box(q.quantize_grouped(&data, &params, gs).unwrap()));
        });

        group.bench_with_input(BenchmarkId::new("dequantize_grouped", name), &n, |b, _| {
            let params = q.calibrate(&data).unwrap();
            let codes = q.quantize_grouped(&data, &params, gs).unwrap();
            b.iter(|| black_box(q.dequantize_grouped(&codes, &params, gs)));
        });
    }
    group.finish();
}

// --- MinMax per-channel (asymmetric) ----------------------------------------

fn bench_per_channel_asymmetric(c: &mut Criterion) {
    let mut group = c.benchmark_group("quant_per_channel_asymmetric");

    // Simulate a weight matrix row x col, quantized per output channel (row).
    let configs: &[(&str, usize, usize, usize)] = &[
        ("768x768_ch0", 768, 768, 0),
        ("4096x1024_ch0", 4096, 1024, 0),
        ("2048x2048_ch1", 2048, 2048, 1),
    ];

    for &(name, rows, cols, axis) in configs {
        let data = make_weights(rows * cols);
        let q = MinMaxQuantizer::new(
            8,
            QuantScheme::Asymmetric,
            QuantGranularity::PerChannel { channel_axis: axis },
        );

        group.bench_with_input(
            BenchmarkId::new("calibrate_2d", name),
            &(rows, cols),
            |b, &(r, c)| {
                b.iter(|| black_box(q.calibrate_2d(&data, r, c).unwrap()));
            },
        );
    }
    group.finish();
}

// --- NF4 encode / decode (QLoRA) --------------------------------------------

fn bench_nf4(c: &mut Criterion) {
    let mut group = c.benchmark_group("quant_nf4");

    let sizes: &[(&str, usize)] = &[
        ("4k", 4_096),
        ("64k", 65_536),
        ("256k", 262_144),
        ("1m", 1_048_576),
    ];

    let qnf4 = Nf4Quantizer::new(64);

    for &(name, n) in sizes {
        let data = make_weights(n);

        group.bench_with_input(BenchmarkId::new("encode", name), &n, |b, _| {
            b.iter(|| black_box(qnf4.encode(&data).unwrap()));
        });

        group.bench_with_input(BenchmarkId::new("decode", name), &n, |b, _| {
            let (packed, absmaxs) = qnf4.encode(&data).unwrap();
            b.iter(|| black_box(qnf4.decode(&packed, &absmaxs).unwrap()));
        });

        group.bench_with_input(BenchmarkId::new("round_trip", name), &n, |b, _| {
            b.iter(|| {
                let (packed, absmaxs) = qnf4.encode(&data).unwrap();
                black_box(qnf4.decode(&packed, &absmaxs).unwrap())
            });
        });

        group.bench_with_input(BenchmarkId::new("mse", name), &n, |b, _| {
            b.iter(|| black_box(qnf4.quantization_mse(&data).unwrap()));
        });
    }
    group.finish();
}

// --- NF4 block sizes ---------------------------------------------------------

fn bench_nf4_block_sizes(c: &mut Criterion) {
    let mut group = c.benchmark_group("nf4_block_size");

    const N: usize = 65_536;
    let data = make_weights(N);

    for bs in [32_usize, 64, 128, 256] {
        let q = Nf4Quantizer::new(bs);
        group.bench_with_input(BenchmarkId::new("encode_64k", bs), &bs, |b, _| {
            b.iter(|| black_box(q.encode(&data).unwrap()));
        });
    }
    group.finish();
}

// --- Magnitude pruning -------------------------------------------------------

fn bench_magnitude_pruning(c: &mut Criterion) {
    let mut group = c.benchmark_group("pruning_magnitude");

    let sizes: &[(&str, usize)] = &[("16k", 16_384), ("256k", 262_144), ("1m", 1_048_576)];

    let sparsities = [0.5_f32, 0.7, 0.9];

    for &(name, n) in sizes {
        let data = make_weights(n);

        for sp in sparsities {
            let label = format!("{name}_sp{}", (sp * 100.0) as u32);

            group.bench_with_input(
                BenchmarkId::new("compute_mask_l1", &label),
                &sp,
                |b, &sp| {
                    let pruner = MagnitudePruner::new(sp, MagnitudeNorm::L1);
                    b.iter(|| black_box(pruner.compute_mask(&data).unwrap()));
                },
            );

            group.bench_with_input(
                BenchmarkId::new("compute_mask_l2", &label),
                &sp,
                |b, &sp| {
                    let pruner = MagnitudePruner::new(sp, MagnitudeNorm::L2);
                    b.iter(|| black_box(pruner.compute_mask(&data).unwrap()));
                },
            );
        }
    }

    // prune in-place
    let sizes_prune: &[(&str, usize)] = &[("256k", 262_144), ("1m", 1_048_576)];
    for &(name, n) in sizes_prune {
        group.bench_with_input(
            BenchmarkId::new("prune_in_place_sp50", name),
            &n,
            |b, &n| {
                b.iter_batched(
                    || make_weights(n),
                    |mut weights| {
                        let pruner = MagnitudePruner::new(0.5, MagnitudeNorm::L1);
                        black_box(pruner.prune(&mut weights).unwrap())
                    },
                    criterion::BatchSize::LargeInput,
                );
            },
        );
    }

    group.finish();
}

// --- criterion wiring --------------------------------------------------------

criterion_group!(
    quant_benches,
    bench_int8_symmetric,
    bench_int4_per_group,
    bench_per_channel_asymmetric,
    bench_nf4,
    bench_nf4_block_sizes,
    bench_magnitude_pruning,
);
criterion_main!(quant_benches);