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

use mircuda::{DenseMatmulPlan, DenseMatmulSpec, DeviceBuffer, PinnedBuffer, f16};

use super::{CudaSequenceScoringModel, scratch::Scratch};
use crate::{
    CudaTensor, Error, Result,
    kernels::{EncoderElementwiseF16, EncoderElementwiseSpec},
};

impl CudaSequenceScoringModel {
    pub fn score(&self, tokens: &[u32]) -> Result<f32> {
        if tokens.is_empty() {
            return Err(Error::InvalidDecoderKernel("empty CUDA scoring pair"));
        }
        if let Some(&token) = tokens.iter().find(|&&token| {
            usize::try_from(token).map_or(true, |token| token >= self.config.vocab_size)
        }) {
            return Err(Error::InvalidToken { token, vocab: self.config.vocab_size });
        }
        let backend = &self.backend;
        let config = &self.config;
        let count = tokens.len();
        let mut ids_host: PinnedBuffer<u32> = backend.inner.context.allocate_pinned(count)?;
        ids_host.copy_from_slice(tokens)?;
        let mut ids = backend.inner.pool.allocate::<u32>(&backend.inner.stream, count)?;
        backend.inner.stream.copy_to_device(&mut ids_host, &mut ids)?;
        let mut scratch =
            Scratch::new(backend, count, config.hidden_size, config.intermediate_size)?;
        ops(backend, count, config.hidden_size, epsilon(config)?)?.embedding_norm(
            &backend.inner.stream,
            &ids,
            f16_tensor(self.tensor("new.embeddings.word_embeddings.weight")?)?,
            f16_tensor(self.tensor("new.embeddings.token_type_embeddings.weight")?)?,
            f16_tensor(self.tensor("new.embeddings.LayerNorm.weight")?)?,
            f16_tensor(self.tensor("new.embeddings.LayerNorm.bias")?)?,
            &mut scratch.first,
        )?;
        for layer in 0..config.num_hidden_layers {
            self.layer(layer, &mut scratch)?;
        }
        let hidden = if config.num_hidden_layers.is_multiple_of(2) {
            &scratch.first
        } else {
            &scratch.second
        };
        backend.inner.stream.copy_device_range(
            hidden,
            0..config.hidden_size,
            &mut scratch.cls,
            0,
        )?;
        project(
            backend,
            1,
            config.hidden_size,
            config.hidden_size,
            &scratch.cls,
            self.tensor("new.pooler.dense.weight")?,
            &mut scratch.pooled_raw,
        )?;
        ops(backend, 1, config.hidden_size, epsilon(config)?)?.tanh_bias(
            &backend.inner.stream,
            &scratch.pooled_raw,
            f16_tensor(self.tensor("new.pooler.dense.bias")?)?,
            &mut scratch.pooled,
        )?;
        project(
            backend,
            1,
            config.hidden_size,
            1,
            &scratch.pooled,
            self.tensor("classifier.weight")?,
            &mut scratch.score_raw,
        )?;
        ops(backend, 1, 1, epsilon(config)?)?.add_bias(
            &backend.inner.stream,
            &scratch.score_raw,
            f16_tensor(self.tensor("classifier.bias")?)?,
            &mut scratch.score,
        )?;
        let mut host: PinnedBuffer<f16> = backend.inner.context.allocate_pinned(1)?;
        backend.inner.stream.copy_to_host(&scratch.score, &mut host)?;
        Ok(host.to_vec()?[0].to_f32())
    }

    pub(super) fn tensor(&self, name: &str) -> Result<&CudaTensor> {
        self.tensors.get(name).ok_or_else(|| Error::MissingTensor(name.into()))
    }
}

pub(super) fn project(
    backend: &crate::CudaBackend,
    rows: usize,
    input_width: usize,
    output_width: usize,
    input: &DeviceBuffer<f16>,
    weight: &CudaTensor,
    output: &mut DeviceBuffer<f16>,
) -> Result<()> {
    let weight = f16_tensor(weight)?;
    let mut plan = DenseMatmulPlan::<f16>::new(
        &backend.inner.context,
        &backend.inner.stream,
        DenseMatmulSpec::new(rows, output_width, input_width)?,
    )?;
    Ok(plan.execute(&backend.inner.stream, input, weight, output, 1.0, 0.0)?)
}

pub(super) fn ops(
    backend: &crate::CudaBackend,
    rows: usize,
    columns: usize,
    epsilon: f32,
) -> Result<EncoderElementwiseF16> {
    EncoderElementwiseF16::compile(
        &backend.inner.compiler,
        EncoderElementwiseSpec { rows, columns, epsilon },
    )
}

pub(super) fn f16_tensor(tensor: &CudaTensor) -> Result<&DeviceBuffer<f16>> {
    tensor.as_f16().ok_or_else(|| Error::DTypeMismatch {
        name: tensor.name().into(),
        expected: "F16",
    })
}

pub(super) fn epsilon(config: &models::layout::EncoderConfig) -> Result<f32> {
    config.layer_norm_eps.to_string().parse().map_err(Into::into)
}

pub(super) fn buffer(backend: &crate::CudaBackend, elements: usize) -> Result<DeviceBuffer<f16>> {
    Ok(backend.inner.pool.allocate::<f16>(&backend.inner.stream, elements)?)
}