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
135
136
137
138
139
140
141
142
143
144

use mircuda::{DeviceBuffer, Stream, bf16};
use models::weights::{BitsAndBytes4BitQuantization, TensorBinding, TensorStorage};

use crate::{
    CudaBackend, CudaTensorSet, Error, Result,
    kernels::{BitsAndBytes4BitLaunch, BitsAndBytes4BitLinear, BitsAndBytes4BitSpec},
};

#[derive(Clone, Debug)]
pub struct BitsAndBytes4BitWeight {
    weight: DeviceBuffer<u8>,
    absmax: DeviceBuffer<u8>,
    quant_map: DeviceBuffer<u8>,
    nested_absmax: DeviceBuffer<u8>,
    nested_quant_map: DeviceBuffer<u8>,
    nested_offset: f32,
    format: BitsAndBytes4BitQuantization,
    input: usize,
    output: usize,
}

impl BitsAndBytes4BitWeight {
    pub(crate) fn load_binding(
        tensors: &CudaTensorSet,
        binding: &TensorBinding,
        input: usize,
        output: usize,
    ) -> Result<Self> {
        let TensorStorage::BitsAndBytes4Bit {
            format,
            absmax,
            quant_map,
            nested_absmax,
            nested_quant_map,
            nested_offset_bits,
            ..
        } = &binding.storage
        else {
            return Err(Error::InvalidQuantizedGemv("binding is not bitsandbytes 4-bit"));
        };
        let weight_tensor = required(tensors, &binding.source)?;
        let absmax_tensor = required(tensors, absmax)?;
        let map_tensor = required(tensors, quant_map)?;
        let nested_absmax_tensor =
            nested_absmax.as_deref().map(|name| required(tensors, name)).transpose()?;
        let nested_map_tensor =
            nested_quant_map.as_deref().map(|name| required(tensors, name)).transpose()?;
        let dummy = absmax_tensor.raw_u8()?;
        let value = Self {
            weight: weight_tensor.raw_u8()?,
            absmax: absmax_tensor.raw_u8()?,
            quant_map: map_tensor.raw_u8()?,
            nested_absmax: nested_absmax_tensor
                .map_or_else(|| Ok(dummy.clone()), crate::CudaTensor::raw_u8)?,
            nested_quant_map: nested_map_tensor
                .map_or_else(|| Ok(dummy), crate::CudaTensor::raw_u8)?,
            nested_offset: nested_offset_bits.map_or(0.0, f32::from_bits),
            format: *format,
            input,
            output,
        };
        value.validate()?;
        Ok(value)
    }

    pub(in crate::backend) fn validate(&self) -> Result<()> {
        let elements = self
            .input
            .checked_mul(self.output)
            .ok_or(Error::InvalidQuantizedGemv("bitsandbytes matrix size overflows"))?;
        require(self.weight.len(), elements.div_ceil(2), "weight")?;
        require(self.quant_map.len(), 16 * 4, "quant map")?;
        let blocks = elements.div_ceil(self.format.block_size);
        if let Some(nested) = self.format.nested_block_size {
            require(self.absmax.len(), blocks, "nested codes")?;
            require(self.nested_absmax.len(), blocks.div_ceil(nested) * 4, "nested absmax")?;
            require(self.nested_quant_map.len(), 256 * 4, "nested quant map")?;
        } else {
            require(self.absmax.len(), blocks * 4, "absmax")?;
        }
        Ok(())
    }
}

#[derive(Clone, Debug)]
pub struct BitsAndBytes4BitBf16Linear {
    operation: BitsAndBytes4BitLinear,
    stream: Stream,
}

impl BitsAndBytes4BitBf16Linear {
    pub(in crate::backend) fn new(
        backend: &CudaBackend,
        tokens: usize,
        weight: &BitsAndBytes4BitWeight,
    ) -> Result<Self> {
        weight.validate()?;
        let spec = BitsAndBytes4BitSpec::new(
            tokens,
            weight.input,
            weight.output,
            weight.format.block_size,
            weight.format.nested_block_size,
        )?;
        Ok(Self {
            operation: BitsAndBytes4BitLinear::compile(&backend.inner.compiler, spec)?,
            stream: backend.inner.stream.clone(),
        })
    }

    pub(in crate::backend) fn execute(
        &self,
        input: &DeviceBuffer<bf16>,
        weight: &BitsAndBytes4BitWeight,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        weight.validate()?;
        self.operation.execute(
            &self.stream,
            &mut BitsAndBytes4BitLaunch {
                input,
                weight: &weight.weight,
                absmax: &weight.absmax,
                quant_map: &weight.quant_map,
                nested_absmax: &weight.nested_absmax,
                nested_quant_map: &weight.nested_quant_map,
                nested_offset: weight.nested_offset,
                output,
            },
        )
    }
}

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

fn require(actual: usize, expected: usize, kind: &'static str) -> Result<()> {
    if actual == expected {
        Ok(())
    } else {
        Err(Error::InvalidQuantizedGemv(kind))
    }
}