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
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

//! Backend parity tests for KvCacheOps.

use super::helpers::*;
use boostr::ops::traits::cache::kv_cache::KvCacheOps;

#[test]
fn test_kv_cache_update_parity() {
    let (cpu_client, cpu_device) = setup_cpu();
    let (b, kv_heads, max_seq, d) = (1, 2, 16, 32);
    let new_len = 4;
    let position = 3;

    // Zero-init caches
    let zeros = vec![0.0f32; b * kv_heads * max_seq * d];
    let k_cache = numr::tensor::Tensor::from_slice(&zeros, &[b, kv_heads, max_seq, d], &cpu_device);
    let v_cache = numr::tensor::Tensor::from_slice(&zeros, &[b, kv_heads, max_seq, d], &cpu_device);
    let new_k = det_tensor(&[b, kv_heads, new_len, d], &cpu_device);
    let new_v = det_tensor(&[b, kv_heads, new_len, d], &cpu_device);

    cpu_client
        .kv_cache_update(&k_cache, &v_cache, &new_k, &new_v, position)
        .unwrap();
    let cpu_k = k_cache.to_vec::<f32>();
    let cpu_v = v_cache.to_vec::<f32>();

    #[cfg(feature = "cuda")]
    with_cuda_backend(|cuda_client, cuda_device| {
        use boostr::ops::traits::cache::kv_cache::KvCacheOps as _;
        use numr::tensor::Tensor;
        let k_c = Tensor::from_slice(
            &vec![0.0f32; b * kv_heads * max_seq * d],
            &[b, kv_heads, max_seq, d],
            &cuda_device,
        );
        let v_c = Tensor::from_slice(
            &vec![0.0f32; b * kv_heads * max_seq * d],
            &[b, kv_heads, max_seq, d],
            &cuda_device,
        );
        let nk = Tensor::from_slice(
            &new_k.to_vec::<f32>(),
            &[b, kv_heads, new_len, d],
            &cuda_device,
        );
        let nv = Tensor::from_slice(
            &new_v.to_vec::<f32>(),
            &[b, kv_heads, new_len, d],
            &cuda_device,
        );
        cuda_client
            .kv_cache_update(&k_c, &v_c, &nk, &nv, position)
            .unwrap();
        assert_parity_f32(
            &k_c.to_vec::<f32>(),
            &cpu_k,
            "kv_cache_update K CUDA vs CPU",
        );
        assert_parity_f32(
            &v_c.to_vec::<f32>(),
            &cpu_v,
            "kv_cache_update V CUDA vs CPU",
        );
    });

    #[cfg(feature = "wgpu")]
    with_wgpu_backend(|wgpu_client, wgpu_device| {
        use boostr::ops::traits::cache::kv_cache::KvCacheOps as _;
        use numr::tensor::Tensor;
        let k_w = Tensor::from_slice(
            &vec![0.0f32; b * kv_heads * max_seq * d],
            &[b, kv_heads, max_seq, d],
            &wgpu_device,
        );
        let v_w = Tensor::from_slice(
            &vec![0.0f32; b * kv_heads * max_seq * d],
            &[b, kv_heads, max_seq, d],
            &wgpu_device,
        );
        let nk = Tensor::from_slice(
            &new_k.to_vec::<f32>(),
            &[b, kv_heads, new_len, d],
            &wgpu_device,
        );
        let nv = Tensor::from_slice(
            &new_v.to_vec::<f32>(),
            &[b, kv_heads, new_len, d],
            &wgpu_device,
        );
        wgpu_client
            .kv_cache_update(&k_w, &v_w, &nk, &nv, position)
            .unwrap();
        assert_parity_f32(
            &k_w.to_vec::<f32>(),
            &cpu_k,
            "kv_cache_update K WGPU vs CPU",
        );
        assert_parity_f32(
            &v_w.to_vec::<f32>(),
            &cpu_v,
            "kv_cache_update V WGPU vs CPU",
        );
    });
}

#[test]
fn test_reshape_and_cache_parity() {
    let (cpu_client, cpu_device) = setup_cpu();
    let num_tokens = 4;
    let num_heads = 2;
    let d = 16;
    let block_size = 4;
    let num_blocks = 2;

    let key = det_tensor(&[num_tokens, num_heads, d], &cpu_device);
    let value = det_tensor(&[num_tokens, num_heads, d], &cpu_device);
    let zeros = vec![0.0f32; num_blocks * block_size * num_heads * d];
    let key_cache = numr::tensor::Tensor::from_slice(
        &zeros,
        &[num_blocks, block_size, num_heads, d],
        &cpu_device,
    );
    let value_cache = numr::tensor::Tensor::from_slice(
        &zeros,
        &[num_blocks, block_size, num_heads, d],
        &cpu_device,
    );
    // Slot mapping: tokens go into slots 0,1,4,5 (block 0 slots 0-1, block 1 slots 0-1)
    let slot_data: Vec<i32> = vec![0, 1, 4, 5];
    let slot_mapping = numr::tensor::Tensor::from_slice(&slot_data, &[num_tokens], &cpu_device);

    cpu_client
        .reshape_and_cache(
            &key,
            &value,
            &key_cache,
            &value_cache,
            &slot_mapping,
            block_size,
        )
        .unwrap();
    let cpu_kc = key_cache.to_vec::<f32>();
    let cpu_vc = value_cache.to_vec::<f32>();

    #[cfg(feature = "cuda")]
    with_cuda_backend(|cuda_client, cuda_device| {
        use boostr::ops::traits::cache::kv_cache::KvCacheOps as _;
        use numr::tensor::Tensor;
        let k = Tensor::from_slice(
            &key.to_vec::<f32>(),
            &[num_tokens, num_heads, d],
            &cuda_device,
        );
        let v = Tensor::from_slice(
            &value.to_vec::<f32>(),
            &[num_tokens, num_heads, d],
            &cuda_device,
        );
        let kc = Tensor::from_slice(
            &vec![0.0f32; num_blocks * block_size * num_heads * d],
            &[num_blocks, block_size, num_heads, d],
            &cuda_device,
        );
        let vc = Tensor::from_slice(
            &vec![0.0f32; num_blocks * block_size * num_heads * d],
            &[num_blocks, block_size, num_heads, d],
            &cuda_device,
        );
        let sm = Tensor::from_slice(&slot_data, &[num_tokens], &cuda_device);
        cuda_client
            .reshape_and_cache(&k, &v, &kc, &vc, &sm, block_size)
            .unwrap();
        assert_parity_f32(
            &kc.to_vec::<f32>(),
            &cpu_kc,
            "reshape_and_cache K CUDA vs CPU",
        );
        assert_parity_f32(
            &vc.to_vec::<f32>(),
            &cpu_vc,
            "reshape_and_cache V CUDA vs CPU",
        );
    });

    #[cfg(feature = "wgpu")]
    with_wgpu_backend(|wgpu_client, wgpu_device| {
        use boostr::ops::traits::cache::kv_cache::KvCacheOps as _;
        use numr::tensor::Tensor;
        let k = Tensor::from_slice(
            &key.to_vec::<f32>(),
            &[num_tokens, num_heads, d],
            &wgpu_device,
        );
        let v = Tensor::from_slice(
            &value.to_vec::<f32>(),
            &[num_tokens, num_heads, d],
            &wgpu_device,
        );
        let kc = Tensor::from_slice(
            &vec![0.0f32; num_blocks * block_size * num_heads * d],
            &[num_blocks, block_size, num_heads, d],
            &wgpu_device,
        );
        let vc = Tensor::from_slice(
            &vec![0.0f32; num_blocks * block_size * num_heads * d],
            &[num_blocks, block_size, num_heads, d],
            &wgpu_device,
        );
        let sm = Tensor::from_slice(&slot_data, &[num_tokens], &wgpu_device);
        wgpu_client
            .reshape_and_cache(&k, &v, &kc, &vc, &sm, block_size)
            .unwrap();
        assert_parity_f32(
            &kc.to_vec::<f32>(),
            &cpu_kc,
            "reshape_and_cache K WGPU vs CPU",
        );
        assert_parity_f32(
            &vc.to_vec::<f32>(),
            &cpu_vc,
            "reshape_and_cache V WGPU vs CPU",
        );
    });
}