libmir-cuda 0.1.0

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

cuda_export!(QuantizeBlockFp8Weight = "libmir_cuda_quantize_block_fp8_weight"(
    source: &DeviceBuffer<bf16>, weight: &mut DeviceBuffer<u8>,
    scales: &mut DeviceBuffer<f32>, rows: u32, columns: u32,
));
cuda_export!(BlockFp8LinearKernel = "libmir_cuda_block_fp8x4_bf16_linear"(
    input: &DeviceBuffer<bf16>, weight: &DeviceBuffer<u8>, scales: &DeviceBuffer<f32>,
    output: &mut DeviceBuffer<bf16>, rows: u32, columns: u32,
));

#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
pub struct BlockFp8LinearSpec {
    pub input_features: usize,
    pub output_features: usize,
}

impl BlockFp8LinearSpec {
    pub fn new(input_features: usize, output_features: usize) -> Result<Self> {
        if input_features == 0
            || output_features == 0
            || !input_features.is_multiple_of(128)
            || !output_features.is_multiple_of(128)
        {
            return Err(Error::InvalidDecoderKernel("invalid block FP8 linear geometry"));
        }
        let _ = product(input_features, output_features)?;
        Ok(Self { input_features, output_features })
    }

    pub fn weight_elements(self) -> Result<usize> {
        product(self.input_features, self.output_features)
    }

    pub fn scale_elements(self) -> Result<usize> {
        product(self.output_features, self.input_features / 128)
    }
}

#[derive(Clone, Debug)]
pub struct BlockFp8LinearKernels {
    quantize: TypedKernel<QuantizeBlockFp8Weight>,
    project: TypedKernel<BlockFp8LinearKernel>,
    spec: BlockFp8LinearSpec,
}

impl BlockFp8LinearKernels {
    pub fn compile(compiler: &Compiler, spec: BlockFp8LinearSpec) -> Result<Self> {
        let source = cuda_kernel_file!("../../kernels/linear_fp8_block.cu");
        let module = compiler.compile(
            source,
            &CompileOptions {
                fast_math: false,
                ..CompileOptions::default()
            },
        )?;
        Ok(Self {
            quantize: module.kernel()?,
            project: module.kernel()?,
            spec,
        })
    }

    pub fn quantize(
        &self,
        stream: &Stream,
        source: &DeviceBuffer<bf16>,
        weight: &mut DeviceBuffer<u8>,
        scales: &mut DeviceBuffer<f32>,
    ) -> Result<()> {
        self.validate_weight(weight, scales)?;
        require("block FP8 source", self.spec.weight_elements()?, source.len())?;
        Ok(self.quantize.launch(
            stream,
            LaunchConfig {
                grid: (
                    narrow(self.spec.output_features)?,
                    narrow(self.spec.input_features / 128)?,
                    1,
                ),
                block: (128, 1, 1),
                shared_memory_bytes: 0,
            },
            (
                source,
                weight,
                scales,
                narrow(self.spec.output_features)?,
                narrow(self.spec.input_features)?,
            ),
        )?)
    }

    pub fn project(
        &self,
        stream: &Stream,
        input: &DeviceBuffer<bf16>,
        weight: &DeviceBuffer<u8>,
        scales: &DeviceBuffer<f32>,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        require("block FP8 input", self.spec.input_features, input.len())?;
        self.validate_weight(weight, scales)?;
        require("block FP8 output", self.spec.output_features, output.len())?;
        Ok(self.project.launch(
            stream,
            LaunchConfig {
                grid: (narrow(self.spec.output_features.div_ceil(64))?, 1, 1),
                block: (256, 1, 1),
                shared_memory_bytes: 0,
            },
            (
                input,
                weight,
                scales,
                output,
                narrow(self.spec.output_features)?,
                narrow(self.spec.input_features)?,
            ),
        )?)
    }

    fn validate_weight(&self, weight: &DeviceBuffer<u8>, scales: &DeviceBuffer<f32>) -> Result<()> {
        require("block FP8 weight", self.spec.weight_elements()?, weight.len())?;
        require("block FP8 scales", self.spec.scale_elements()?, scales.len())
    }
}