libmir-cuda 0.2.0

CUDA inference backend for libmir
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

use mircuda::{DeviceBuffer, DeviceElement, bf16};

use super::super::CudaBackend;
use crate::{
    CudaConfig, Result,
    kernels::{
        SpatialMergeKernels, VisionAttention, VisionAttentionSpec, VisionClip, VisionClipSpec,
        VisionElementwise, VisionElementwiseSpec, VisionEmbeddingSplice, VisionPool,
        VisionPoolSpec, VisionSpatialRope,
    },
};

#[test]
fn splices_image_embeddings_without_host_roundtrip()
-> std::result::Result<(), Box<dyn std::error::Error>> {
    let backend = CudaBackend::new(CudaConfig::default())?;
    let image = copy(&backend, &bf16s(&[10.0, 11.0, 20.0, 21.0]))?;
    let mut hidden = copy(&backend, &bf16s(&[1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0]))?;
    VisionEmbeddingSplice::compile(&backend.inner.compiler, 2, 2, 1)?.execute(
        &backend.inner.stream,
        &image,
        &mut hidden,
    )?;
    assert_eq!(
        read(&backend, &hidden)?.iter().map(|value| value.to_f32()).collect::<Vec<_>>(),
        [1.0, 2.0, 10.0, 11.0, 20.0, 21.0, 7.0, 8.0]
    );
    Ok(())
}

#[test]
fn pools_and_standardizes_in_the_vision_kernel_module()
-> std::result::Result<(), Box<dyn std::error::Error>> {
    let backend = CudaBackend::new(CudaConfig::default())?;
    let input = copy(&backend, &bf16s(&[1.0, 2.0, 3.0, 4.0]))?;
    let bias = copy(&backend, &bf16s(&[0.0]))?;
    let scale = copy(&backend, &bf16s(&[2.0]))?;
    let mut pooled = backend.inner.pool.allocate(&backend.inner.stream, 1)?;
    let mut output = backend.inner.pool.allocate(&backend.inner.stream, 1)?;
    let operation = VisionPool::compile(
        &backend.inner.compiler,
        VisionPoolSpec {
            grid_height: 2,
            grid_width: 2,
            hidden: 1,
            kernel: 2,
        },
    )?;
    operation.execute(&backend.inner.stream, &input, &mut pooled)?;
    operation.standardize(&backend.inner.stream, &pooled, &bias, &scale, &mut output)?;
    assert!((read(&backend, &output)?[0].to_f32() - 5.0).abs() < 0.05);
    Ok(())
}

#[test]
fn applies_vision_bias_in_place_without_a_second_device_buffer()
-> std::result::Result<(), Box<dyn std::error::Error>> {
    let backend = CudaBackend::new(CudaConfig::default())?;
    let mut values = copy(&backend, &bf16s(&[1.0, 2.0, 3.0, 4.0]))?;
    let input = values.clone();
    let bias = copy(&backend, &bf16s(&[10.0, 20.0]))?;
    VisionElementwise::compile(
        &backend.inner.compiler,
        VisionElementwiseSpec { rows: 2, columns: 2, epsilon: 0.0 },
    )?
    .add_bias(&backend.inner.stream, &input, &bias, &mut values)?;
    assert_eq!(as_f32(&read(&backend, &values)?), [11.0, 22.0, 13.0, 24.0]);
    Ok(())
}

#[test]
fn compiles_the_complete_vision_primitive_set()
-> std::result::Result<(), Box<dyn std::error::Error>> {
    let backend = CudaBackend::new(CudaConfig::default())?;
    let compiler = &backend.inner.compiler;
    VisionElementwise::compile(
        compiler,
        VisionElementwiseSpec { rows: 2, columns: 4, epsilon: 1.0e-6 },
    )?;
    VisionClip::compile(compiler, VisionClipSpec { rows: 2, columns: 4 })?;
    let attention = VisionAttention::compile(
        compiler,
        VisionAttentionSpec {
            tokens: 2,
            query_heads: 2,
            kv_heads: 1,
            head_dim: 4,
            scale: 1.0,
        },
    )?;
    VisionSpatialRope::compile(compiler, 2, 2, 4, 10_000.0)?;
    SpatialMergeKernels::compile(compiler)?;
    let query = copy(&backend, &bf16s(&[0.0; 16]))?;
    let key = copy(&backend, &bf16s(&[0.0; 8]))?;
    let value = copy(&backend, &bf16s(&[1.0, 2.0, 3.0, 4.0, 3.0, 4.0, 5.0, 6.0]))?;
    let mut output = backend.inner.pool.allocate(&backend.inner.stream, 16)?;
    attention.execute(&backend.inner.stream, &query, &key, &value, &mut output)?;
    let actual = read(&backend, &output)?;
    for row in actual.as_chunks::<4>().0 {
        for (actual, expected) in row.iter().zip([2.0, 3.0, 4.0, 5.0]) {
            assert!((actual.to_f32() - expected).abs() < 0.05);
        }
    }
    Ok(())
}

#[test]
fn executes_spatial_merge_interpolation_split_and_rope()
-> std::result::Result<(), Box<dyn std::error::Error>> {
    let backend = CudaBackend::new(CudaConfig::default())?;
    let kernels = SpatialMergeKernels::compile(&backend.inner.compiler)?;
    let table = copy(&backend, &bf16s(&[0.0, 10.0, 20.0, 30.0]))?;
    let mut interpolated = backend.inner.pool.allocate(&backend.inner.stream, 9)?;
    kernels.interpolate(&backend.inner.stream, &table, &mut interpolated, 3, 3, 2, 1, 1)?;
    let actual = read(&backend, &interpolated)?
        .iter()
        .map(|value| value.to_f32())
        .collect::<Vec<_>>();
    assert_eq!(actual, [0.0, 5.0, 10.0, 10.0, 15.0, 20.0, 20.0, 25.0, 30.0]);

    let qkv_values = [
        1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0,
        17.0, 18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0,
    ];
    let qkv = copy(&backend, &bf16s(&qkv_values))?;
    let mut query = backend.inner.pool.allocate(&backend.inner.stream, 8)?;
    let mut key = backend.inner.pool.allocate(&backend.inner.stream, 8)?;
    let mut value = backend.inner.pool.allocate(&backend.inner.stream, 8)?;
    kernels.split_qkv(&backend.inner.stream, &qkv, &mut query, &mut key, &mut value, 1, 8)?;
    assert_eq!(as_f32(&read(&backend, &query)?), qkv_values[..8]);
    assert_eq!(as_f32(&read(&backend, &key)?), qkv_values[8..16]);
    assert_eq!(as_f32(&read(&backend, &value)?), qkv_values[16..]);

    let positions = copy(&backend, &[1_u32, 2])?;
    let mut rotated = backend.inner.pool.allocate(&backend.inner.stream, 8)?;
    kernels.rope(&backend.inner.stream, &query, &positions, &mut rotated, 1, 1, 8)?;
    let angles = [1.0_f32, 0.01, 2.0, 0.02, 1.0, 0.01, 2.0, 0.02];
    let paired = [-5.0_f32, -6.0, -7.0, -8.0, 1.0, 2.0, 3.0, 4.0];
    for (((actual, input), paired), angle) in read(&backend, &rotated)?
        .iter()
        .zip(qkv_values[..8].iter().copied())
        .zip(paired)
        .zip(angles)
    {
        let expected = paired.mul_add(angle.sin(), input * angle.cos());
        assert!((actual.to_f32() - expected).abs() < 0.05);
    }
    Ok(())
}

fn bf16s(values: &[f32]) -> Vec<bf16> {
    values.iter().copied().map(bf16::from_f32).collect()
}

fn as_f32(values: &[bf16]) -> Vec<f32> {
    values.iter().map(|value| value.to_f32()).collect()
}

fn copy<T: DeviceElement>(backend: &CudaBackend, values: &[T]) -> Result<DeviceBuffer<T>> {
    let mut host = backend.inner.context.allocate_pinned(values.len())?;
    host.copy_from_slice(values)?;
    let mut device = backend.inner.pool.allocate(&backend.inner.stream, values.len())?;
    backend.inner.stream.copy_to_device(&mut host, &mut device)?;
    Ok(device)
}

fn read<T: DeviceElement>(backend: &CudaBackend, source: &DeviceBuffer<T>) -> Result<Vec<T>> {
    let mut host = backend.inner.context.allocate_pinned(source.len())?;
    backend.inner.stream.copy_to_host(source, &mut host)?;
    Ok(host.to_vec()?)
}