libmir-cuda 0.3.0

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

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!(PackedInt8EmbeddingKernel = "libmir_cuda_packed_int8_embedding_bf16"(
    selected: &DeviceBuffer<u32>, weight: &DeviceBuffer<i32>, scales: &DeviceBuffer<bf16>,
    output: &mut DeviceBuffer<bf16>, selected_start: u32, tokens: u32, vocab: u32, hidden: u32,
    output_scale: f32, bits: u32, group_size: u32,
));

#[derive(Clone, Copy, Debug)]
pub struct PackedInt8EmbeddingSpec {
    pub vocab: usize,
    pub hidden: usize,
    pub output_scale: f32,
    bits: usize,
    group_size: usize,
}

impl PackedInt8EmbeddingSpec {
    pub const fn new(vocab: usize, hidden: usize, output_scale: f32) -> Result<Self> {
        Self::new_packed(vocab, hidden, output_scale, 8, hidden)
    }

    pub const fn new_packed(
        vocab: usize,
        hidden: usize,
        output_scale: f32,
        bits: usize,
        group_size: usize,
    ) -> Result<Self> {
        if vocab == 0
            || hidden == 0
            || !matches!(bits, 4 | 8)
            || group_size == 0
            || !hidden.is_multiple_of(group_size)
            || !(hidden * bits).is_multiple_of(32)
        {
            return Err(Error::InvalidQuantizedGemv(
                "packed integer embedding has invalid dimensions, bits, or group size",
            ));
        }
        if !output_scale.is_finite() {
            return Err(Error::InvalidQuantizedGemv(
                "packed integer embedding scale must be finite",
            ));
        }
        Ok(Self {
            vocab,
            hidden,
            output_scale,
            bits,
            group_size,
        })
    }
}

pub struct PackedInt8EmbeddingLaunch<'a> {
    pub selected: &'a DeviceBuffer<u32>,
    pub selected_start: usize,
    pub tokens: usize,
    pub weight: &'a DeviceBuffer<i32>,
    pub scales: &'a DeviceBuffer<bf16>,
    pub output: &'a mut DeviceBuffer<bf16>,
}

#[derive(Clone, Debug)]
pub struct PackedInt8Embedding {
    kernel: TypedKernel<PackedInt8EmbeddingKernel>,
    spec: PackedInt8EmbeddingSpec,
}

impl PackedInt8Embedding {
    pub fn compile(compiler: &Compiler, spec: PackedInt8EmbeddingSpec) -> Result<Self> {
        let source = cuda_kernel_file!("../../../kernels/packed_int8_bf16.cu");
        let module =
            compiler.compile(source, &CompileOptions { fast_math: true, ..Default::default() })?;
        Ok(Self { kernel: module.kernel()?, spec })
    }

    pub fn execute(
        &self,
        stream: &Stream,
        launch: &mut PackedInt8EmbeddingLaunch<'_>,
    ) -> Result<()> {
        let selected_end = launch
            .selected_start
            .checked_add(launch.tokens)
            .ok_or(Error::InvalidQuantizedGemv("packed integer selected range overflow"))?;
        require("packed integer selected tokens", selected_end, launch.selected.len())?;
        require(
            "packed integer embedding weight",
            product(self.spec.vocab, self.spec.hidden * self.spec.bits / 32)?,
            launch.weight.len(),
        )?;
        require(
            "packed integer embedding scales",
            product(self.spec.vocab, self.spec.hidden / self.spec.group_size)?,
            launch.scales.len(),
        )?;
        require(
            "packed integer embedding output",
            product(launch.tokens, self.spec.hidden)?,
            launch.output.len(),
        )?;
        Ok(self.kernel.launch(
            stream,
            LaunchConfig {
                grid: (narrow(self.spec.hidden.div_ceil(256))?, narrow(launch.tokens)?, 1),
                block: (256, 1, 1),
                shared_memory_bytes: 0,
            },
            (
                launch.selected,
                launch.weight,
                launch.scales,
                &mut *launch.output,
                narrow(launch.selected_start)?,
                narrow(launch.tokens)?,
                narrow(self.spec.vocab)?,
                narrow(self.spec.hidden)?,
                self.spec.output_scale,
                narrow(self.spec.bits)?,
                narrow(self.spec.group_size)?,
            ),
        )?)
    }
}