boostr 0.1.0

ML framework built on numr - attention, quantization, model architectures
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

//! Backend parity tests for KvCacheQuantOps.

use super::helpers::*;
use boostr::ops::traits::cache::kv_cache_quant::{Int4GroupSize, KvCacheQuantOps};

#[test]
fn test_quantize_dequantize_fp8_roundtrip_parity() {
    let (cpu_client, cpu_device) = setup_cpu();
    let num_tokens = 8;
    let head_dim = 32;
    let input = det_tensor(&[num_tokens, head_dim], &cpu_device);

    let (quantized, scales) = cpu_client
        .quantize_kv_fp8_per_token(&input, num_tokens, head_dim)
        .unwrap();
    let cpu_deq = cpu_client
        .dequantize_kv_fp8_per_token(
            &quantized,
            &scales,
            num_tokens,
            head_dim,
            numr::dtype::DType::F32,
        )
        .unwrap();
    let cpu_deq_vec = cpu_deq.to_vec::<f32>();

    #[cfg(feature = "cuda")]
    with_cuda_backend(|cuda_client, cuda_device| {
        use boostr::ops::traits::cache::kv_cache_quant::KvCacheQuantOps as _;
        use numr::tensor::Tensor;
        let inp = Tensor::from_slice(
            &input.to_vec::<f32>(),
            &[num_tokens, head_dim],
            &cuda_device,
        );
        let (q, s) = cuda_client
            .quantize_kv_fp8_per_token(&inp, num_tokens, head_dim)
            .unwrap();
        let deq = cuda_client
            .dequantize_kv_fp8_per_token(&q, &s, num_tokens, head_dim, numr::dtype::DType::F32)
            .unwrap();
        // FP8 quantization is inherently lossy; CUDA path goes F32→F16→FP8→F16→F32
        // while CPU does F32→FP8→F32 directly, so wider tolerance needed
        assert_parity_f32_tol(
            &deq.to_vec::<f32>(),
            &cpu_deq_vec,
            "fp8 roundtrip CUDA vs CPU",
            0.1,  // 10% relative — FP8 has only 3 mantissa bits
            0.01, // absolute tolerance for values near zero
        );
    });

    #[cfg(feature = "wgpu")]
    with_wgpu_backend(|wgpu_client, wgpu_device| {
        use boostr::ops::traits::cache::kv_cache_quant::KvCacheQuantOps as _;
        use numr::tensor::Tensor;
        let inp = Tensor::from_slice(
            &input.to_vec::<f32>(),
            &[num_tokens, head_dim],
            &wgpu_device,
        );
        let (q, s) = wgpu_client
            .quantize_kv_fp8_per_token(&inp, num_tokens, head_dim)
            .unwrap();
        let deq = wgpu_client
            .dequantize_kv_fp8_per_token(&q, &s, num_tokens, head_dim, numr::dtype::DType::F32)
            .unwrap();
        assert_parity_f32(
            &deq.to_vec::<f32>(),
            &cpu_deq_vec,
            "fp8 roundtrip WGPU vs CPU",
        );
    });
}

#[test]
fn test_quantize_dequantize_int4_roundtrip_parity() {
    let (cpu_client, cpu_device) = setup_cpu();
    let num_tokens = 8;
    let head_dim = 64;
    let group_size = Int4GroupSize::Group64;
    let input = det_tensor(&[num_tokens, head_dim], &cpu_device);

    let (packed, scales, zeros) = cpu_client
        .quantize_kv_int4(&input, num_tokens, head_dim, group_size)
        .unwrap();
    let cpu_deq = cpu_client
        .dequantize_kv_int4(&packed, &scales, &zeros, num_tokens, head_dim, group_size)
        .unwrap();
    let cpu_deq_vec = cpu_deq.to_vec::<f32>();

    #[cfg(feature = "cuda")]
    with_cuda_backend(|cuda_client, cuda_device| {
        use boostr::ops::traits::cache::kv_cache_quant::KvCacheQuantOps as _;
        use numr::tensor::Tensor;
        let inp = Tensor::from_slice(
            &input.to_vec::<f32>(),
            &[num_tokens, head_dim],
            &cuda_device,
        );
        let (p, s, z) = cuda_client
            .quantize_kv_int4(&inp, num_tokens, head_dim, group_size)
            .unwrap();
        let deq = cuda_client
            .dequantize_kv_int4(&p, &s, &z, num_tokens, head_dim, group_size)
            .unwrap();
        assert_parity_f32(
            &deq.to_vec::<f32>(),
            &cpu_deq_vec,
            "int4 roundtrip CUDA vs CPU",
        );
    });

    #[cfg(feature = "wgpu")]
    with_wgpu_backend(|wgpu_client, wgpu_device| {
        use boostr::ops::traits::cache::kv_cache_quant::KvCacheQuantOps as _;
        use numr::tensor::Tensor;
        let inp = Tensor::from_slice(
            &input.to_vec::<f32>(),
            &[num_tokens, head_dim],
            &wgpu_device,
        );
        let (p, s, z) = wgpu_client
            .quantize_kv_int4(&inp, num_tokens, head_dim, group_size)
            .unwrap();
        let deq = wgpu_client
            .dequantize_kv_int4(&p, &s, &z, num_tokens, head_dim, group_size)
            .unwrap();
        assert_parity_f32(
            &deq.to_vec::<f32>(),
            &cpu_deq_vec,
            "int4 roundtrip WGPU vs CPU",
        );
    });
}

#[test]
fn test_quantize_dequantize_int8_roundtrip_parity() {
    let (cpu_client, cpu_device) = setup_cpu();
    let num_tokens = 8;
    let head_dim = 32;
    let input = det_tensor(&[num_tokens, head_dim], &cpu_device);

    let (quantized, scales) = cpu_client
        .quantize_kv_int8(&input, num_tokens, head_dim)
        .unwrap();
    let cpu_deq = cpu_client
        .dequantize_kv_int8(&quantized, &scales, num_tokens, head_dim)
        .unwrap();
    let cpu_deq_vec = cpu_deq.to_vec::<f32>();

    #[cfg(feature = "cuda")]
    with_cuda_backend(|cuda_client, cuda_device| {
        use boostr::ops::traits::cache::kv_cache_quant::KvCacheQuantOps as _;
        use numr::tensor::Tensor;
        let inp = Tensor::from_slice(
            &input.to_vec::<f32>(),
            &[num_tokens, head_dim],
            &cuda_device,
        );
        let (q, s) = cuda_client
            .quantize_kv_int8(&inp, num_tokens, head_dim)
            .unwrap();
        let deq = cuda_client
            .dequantize_kv_int8(&q, &s, num_tokens, head_dim)
            .unwrap();
        assert_parity_f32(
            &deq.to_vec::<f32>(),
            &cpu_deq_vec,
            "int8 roundtrip CUDA vs CPU",
        );
    });

    #[cfg(feature = "wgpu")]
    with_wgpu_backend(|wgpu_client, wgpu_device| {
        use boostr::ops::traits::cache::kv_cache_quant::KvCacheQuantOps as _;
        use numr::tensor::Tensor;
        let inp = Tensor::from_slice(
            &input.to_vec::<f32>(),
            &[num_tokens, head_dim],
            &wgpu_device,
        );
        let (q, s) = wgpu_client
            .quantize_kv_int8(&inp, num_tokens, head_dim)
            .unwrap();
        let deq = wgpu_client
            .dequantize_kv_int8(&q, &s, num_tokens, head_dim)
            .unwrap();
        assert_parity_f32(
            &deq.to_vec::<f32>(),
            &cpu_deq_vec,
            "int8 roundtrip WGPU vs CPU",
        );
    });
}