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
134
135
136
137
138
139
140
141

use models::weights::{TensorBinding, TensorStorage};

use crate::{
    AffineQuantizedConfig, AffineQuantizedTensors, CudaTensor, CudaTensorDType, CudaTensorSet,
    Error, Result,
};

#[derive(Clone, Debug)]
pub struct AffineQuantizedWeight {
    pub weight: CudaTensor,
    pub scales: CudaTensor,
    pub biases: CudaTensor,
}

impl AffineQuantizedWeight {
    pub fn load(tensors: &CudaTensorSet, prefix: &str) -> Result<Self> {
        Ok(Self {
            weight: required(tensors, &format!("{prefix}.weight"))?,
            scales: required(tensors, &format!("{prefix}.scales"))?,
            biases: required(tensors, &format!("{prefix}.biases"))?,
        })
    }

    pub(crate) fn load_binding(tensors: &CudaTensorSet, binding: &TensorBinding) -> Result<Self> {
        let TensorStorage::AffineQuantized { scales, biases: Some(biases), .. } = &binding.storage
        else {
            return Err(Error::InvalidQuantizedGemv(
                "binding is not an affine weight with zero-point biases",
            ));
        };
        Ok(Self {
            weight: required(tensors, &binding.source)?,
            scales: required(tensors, scales)?,
            biases: required(tensors, biases)?,
        })
    }

    pub(crate) fn validate(
        &self,
        matrices: usize,
        input: usize,
        output: usize,
        group_size: usize,
        bits: usize,
    ) -> Result<()> {
        let values_per_word = 32_usize
            .checked_div(bits)
            .filter(|value| *value != 0)
            .ok_or(Error::InvalidQuantizedGemv("invalid affine precision"))?;
        let packed = input / values_per_word;
        let groups = input / group_size;
        shape(&self.weight, expected(matrices, output, packed))?;
        shape(&self.scales, expected(matrices, output, groups))?;
        shape(&self.biases, expected(matrices, output, groups))?;
        dtype(&self.weight, CudaTensorDType::U32, "U32")?;
        dtype(&self.scales, CudaTensorDType::Bf16, "BF16")?;
        dtype(&self.biases, CudaTensorDType::Bf16, "BF16")
    }

    /// Derives affine precision and group geometry from tensor shapes.
    pub fn infer_config(
        &self,
        matrices: usize,
        input: usize,
        output: usize,
    ) -> Result<AffineQuantizedConfig> {
        let packed = trailing(&self.weight, matrices, output)?;
        let groups = trailing(&self.scales, matrices, output)?;
        if trailing(&self.biases, matrices, output)? != groups || groups == 0 || input == 0 {
            return Err(Error::InvalidQuantizedGemv("inconsistent affine group tensors"));
        }
        let packed_bits = packed
            .checked_mul(32)
            .ok_or(Error::InvalidQuantizedGemv("affine packed width overflow"))?;
        if !packed_bits.is_multiple_of(input) || !input.is_multiple_of(groups) {
            return Err(Error::InvalidQuantizedGemv("affine tensor shape is not integral"));
        }
        let config = AffineQuantizedConfig::new(input, output, input / groups, packed_bits / input);
        self.validate(matrices, input, output, config.group_size, config.bits)?;
        Ok(config)
    }

    pub(crate) const fn tensors(&self) -> AffineQuantizedTensors<'_> {
        AffineQuantizedTensors {
            weight: &self.weight,
            scales: &self.scales,
            biases: &self.biases,
        }
    }
}

fn trailing(tensor: &CudaTensor, matrices: usize, output: usize) -> Result<usize> {
    let shape = tensor.shape();
    let value = match (matrices, shape) {
        (1, [rows, trailing]) if *rows == output => Some(*trailing),
        (_, [banks, rows, trailing]) if *banks == matrices && *rows == output => Some(*trailing),
        _ => None,
    };
    value.ok_or_else(|| Error::InvalidQuantizedTensor {
        name: tensor.name().into(),
        expected: expected(matrices, output, 0),
        actual: shape.to_vec(),
    })
}

fn expected(matrices: usize, output: usize, trailing: usize) -> Vec<usize> {
    if matrices == 1 {
        vec![output, trailing]
    } else {
        vec![matrices, output, trailing]
    }
}

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

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

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