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

use mircuda::{
    CompileOptions, Compiler, DeviceBuffer, LaunchConfig, Stream, TypedKernel, bf16, cuda_export,
    cuda_kernel_file,
};

use super::super::geometry::{narrow, product, require};
use crate::{Error, Result};

cuda_export!(
    Int4Kernel = "libmir_cuda_affine_embedding_bf16_int4"(
        weight: &DeviceBuffer<u32>, scales: &DeviceBuffer<bf16>, biases: &DeviceBuffer<bf16>,
        selected: &DeviceBuffer<u32>, output: &mut DeviceBuffer<bf16>, selected_start: u32,
        tokens: u32, vocab: u32, hidden: u32, group_size: u32, output_scale: f32,
    )
);
cuda_export!(
    Int8Kernel = "libmir_cuda_affine_embedding_bf16_int8"(
        weight: &DeviceBuffer<u32>, scales: &DeviceBuffer<bf16>, biases: &DeviceBuffer<bf16>,
        selected: &DeviceBuffer<u32>, output: &mut DeviceBuffer<bf16>, selected_start: u32,
        tokens: u32, vocab: u32, hidden: u32, group_size: u32, output_scale: f32,
    )
);

#[derive(Clone, Copy, Debug, PartialEq)]
pub struct AffineEmbeddingSpec {
    pub vocab: usize,
    pub hidden: usize,
    pub group_size: usize,
    pub bits: usize,
    pub output_scale: f32,
}

#[derive(Clone, Debug)]
pub struct AffineEmbedding {
    kernel: Kernel,
    spec: AffineEmbeddingSpec,
}

#[derive(Clone, Debug)]
enum Kernel {
    Int4(TypedKernel<Int4Kernel>),
    Int8(TypedKernel<Int8Kernel>),
}

impl AffineEmbedding {
    pub fn compile(compiler: &Compiler, spec: AffineEmbeddingSpec) -> Result<Self> {
        validate(spec)?;
        let source = cuda_kernel_file!("../../../kernels/affine_embedding_bf16.cu");
        let module = compiler.compile(source, &CompileOptions::default())?;
        let kernel = match spec.bits {
            4 => Kernel::Int4(module.kernel()?),
            8 => Kernel::Int8(module.kernel()?),
            _ => return Err(Error::InvalidQuantizedGemv("unsupported embedding precision")),
        };
        Ok(Self { kernel, spec })
    }

    #[allow(clippy::too_many_arguments)]
    pub fn execute(
        &self,
        stream: &Stream,
        weight: &DeviceBuffer<u32>,
        scales: &DeviceBuffer<bf16>,
        biases: &DeviceBuffer<bf16>,
        selected: &DeviceBuffer<u32>,
        selected_start: usize,
        tokens: usize,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        if tokens == 0 {
            return Err(Error::InvalidQuantizedGemv("embedding token batch is empty"));
        }
        let values_per_word = 32 / self.spec.bits;
        require(
            "affine embedding weight",
            product(self.spec.vocab, self.spec.hidden / values_per_word)?,
            weight.len(),
        )?;
        let groups = product(self.spec.vocab, self.spec.hidden / self.spec.group_size)?;
        require("affine embedding scales", groups, scales.len())?;
        require("affine embedding biases", groups, biases.len())?;
        require(
            "affine embedding selected",
            selected_start
                .checked_add(tokens)
                .ok_or(Error::InvalidQuantizedGemv("embedding token index overflow"))?,
            selected.len(),
        )?;
        let elements = product(tokens, self.spec.hidden)?;
        require("affine embedding output", elements, output.len())?;
        let launch = LaunchConfig {
            grid: (narrow(elements.div_ceil(256))?, 1, 1),
            block: (256, 1, 1),
            shared_memory_bytes: 0,
        };
        let arguments = (
            weight,
            scales,
            biases,
            selected,
            output,
            narrow(selected_start)?,
            narrow(tokens)?,
            narrow(self.spec.vocab)?,
            narrow(self.spec.hidden)?,
            narrow(self.spec.group_size)?,
            self.spec.output_scale,
        );
        Ok(match &self.kernel {
            Kernel::Int4(kernel) => kernel.launch(stream, launch, arguments),
            Kernel::Int8(kernel) => kernel.launch(stream, launch, arguments),
        }?)
    }
}

fn validate(spec: AffineEmbeddingSpec) -> Result<()> {
    let values_per_word = match spec.bits {
        4 => 8,
        8 => 4,
        _ => return Err(Error::InvalidQuantizedGemv("unsupported embedding precision")),
    };
    if spec.vocab == 0
        || spec.hidden == 0
        || spec.group_size == 0
        || !spec.hidden.is_multiple_of(spec.group_size)
        || !spec.hidden.is_multiple_of(values_per_word)
        || !spec.output_scale.is_finite()
        || spec.output_scale <= 0.0
    {
        return Err(Error::InvalidQuantizedGemv("invalid affine embedding geometry"));
    }
    Ok(())
}