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
133
134

use mircuda::{DeviceBuffer, Stream, bf16};
use models::weights::{TensorBinding, TensorStorage};

use crate::{
    CudaBackend, CudaTensor, CudaTensorDType, CudaTensorSet, Error, Result,
    kernels::{AwqLaunch, AwqLinear, AwqSpec},
};

#[derive(Clone, Debug)]
pub struct AwqWeight {
    weight: CudaTensor,
    zero_points: CudaTensor,
    scales: CudaTensor,
    group_size: usize,
}

impl AwqWeight {
    pub(crate) fn load_binding(
        tensors: &CudaTensorSet,
        binding: &TensorBinding,
        input: usize,
        output: usize,
    ) -> Result<Self> {
        let TensorStorage::Awq { format, scales, zero_points } = &binding.storage else {
            return Err(Error::InvalidQuantizedGemv("binding is not an AWQ weight"));
        };
        if !format.is_gemm_w4a16() {
            return Err(Error::InvalidQuantizedGemv("CUDA requires AWQ GEMM W4A16 storage"));
        }
        let value = Self {
            weight: required(tensors, &binding.source)?,
            zero_points: required(tensors, zero_points)?,
            scales: required(tensors, scales)?,
            group_size: format.group_size,
        };
        value.validate(input, output)?;
        Ok(value)
    }

    pub(in crate::backend) fn validate(&self, input: usize, output: usize) -> Result<()> {
        if self.group_size == 0 || !input.is_multiple_of(self.group_size) {
            return Err(Error::InvalidQuantizedGemv("AWQ group geometry is invalid"));
        }
        shape(&self.weight, &[input, output / 8])?;
        shape(&self.zero_points, &[input / self.group_size, output / 8])?;
        shape(&self.scales, &[input / self.group_size, output])?;
        dtype(&self.weight, CudaTensorDType::I32, "I32")?;
        dtype(&self.zero_points, CudaTensorDType::I32, "I32")?;
        dtype(&self.scales, CudaTensorDType::F16, "F16")
    }
}

#[derive(Clone, Debug)]
pub struct AwqBf16Linear {
    operation: AwqLinear,
    stream: Stream,
    input: usize,
    output: usize,
}

impl AwqBf16Linear {
    pub(in crate::backend) fn new(
        backend: &CudaBackend,
        tokens: usize,
        input: usize,
        output: usize,
        weight: &AwqWeight,
    ) -> Result<Self> {
        weight.validate(input, output)?;
        Ok(Self {
            operation: AwqLinear::compile(
                &backend.inner.compiler,
                AwqSpec::new(tokens, input, output, weight.group_size)?,
            )?,
            stream: backend.inner.stream.clone(),
            input,
            output,
        })
    }

    pub(in crate::backend) fn execute(
        &self,
        input: &DeviceBuffer<bf16>,
        weight: &AwqWeight,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        weight.validate(self.input, self.output)?;
        self.operation.execute(
            &self.stream,
            &mut AwqLaunch {
                input,
                weight: i32_buffer(&weight.weight)?,
                zero_points: i32_buffer(&weight.zero_points)?,
                scales: weight.scales.as_f16().ok_or_else(|| Error::DTypeMismatch {
                    name: weight.scales.name().into(),
                    expected: "F16",
                })?,
                output,
            },
        )
    }
}

fn required(tensors: &CudaTensorSet, name: &str) -> Result<CudaTensor> {
    tensors.get(name).cloned().ok_or_else(|| Error::MissingTensor(name.into()))
}

fn i32_buffer(tensor: &CudaTensor) -> Result<&DeviceBuffer<i32>> {
    tensor.as_i32().ok_or_else(|| Error::DTypeMismatch {
        name: tensor.name().into(),
        expected: "I32",
    })
}

fn shape(tensor: &CudaTensor, expected: &[usize]) -> Result<()> {
    if tensor.shape() != expected {
        return Err(Error::InvalidQuantizedTensor {
            name: tensor.name().into(),
            expected: expected.to_vec(),
            actual: tensor.shape().to_vec(),
        });
    }
    Ok(())
}

fn dtype(tensor: &CudaTensor, expected: CudaTensorDType, name: &'static str) -> Result<()> {
    if tensor.dtype() != expected {
        return Err(Error::DTypeMismatch {
            name: tensor.name().into(),
            expected: name,
        });
    }
    Ok(())
}