libmir-metal 0.2.0

Metal 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
145
146
147

use crate::engine::{Array, Error, ModelTensors, Result, Stream};

#[derive(Debug)]
pub struct DenseLinear {
    transposed_weight: Array,
    bias: Option<Array>,
    clipping: Option<LinearClipping>,
}

#[derive(Debug)]
struct LinearClipping {
    input_minimum: Array,
    input_maximum: Array,
    output_minimum: Array,
    output_maximum: Array,
}

impl DenseLinear {
    pub fn load(tensors: &ModelTensors, prefix: &str, stream: &Stream) -> Result<Self> {
        Self::load_names(
            tensors,
            &format!("{prefix}.weight"),
            Some(&format!("{prefix}.bias")),
            None,
            stream,
        )
    }

    pub fn load_clippable(tensors: &ModelTensors, prefix: &str, stream: &Stream) -> Result<Self> {
        let linear = format!("{prefix}.linear");
        if tensors.contains(&format!("{linear}.weight"))? {
            Self::load_names(
                tensors,
                &format!("{linear}.weight"),
                Some(&format!("{linear}.bias")),
                Some(prefix),
                stream,
            )
        } else {
            Self::load_names(
                tensors,
                &format!("{prefix}.weight"),
                Some(&format!("{prefix}.bias")),
                Some(prefix),
                stream,
            )
        }
    }

    pub fn load_flattened(tensors: &ModelTensors, prefix: &str, stream: &Stream) -> Result<Self> {
        let weight = tensors.get(&format!("{prefix}.weight"))?;
        let shape = weight.shape()?;
        let output = *shape.first().ok_or(Error::ShapeOverflow)?;
        let input = shape[1..].iter().try_fold(1_i32, |total, dimension| {
            total.checked_mul(*dimension).ok_or(Error::ShapeOverflow)
        })?;
        let weight = weight.reshape(&[output, input], stream)?;
        Ok(Self {
            transposed_weight: weight.transpose(&[1, 0], stream)?,
            bias: tensors.get_optional(&format!("{prefix}.bias"))?,
            clipping: None,
        })
    }

    pub(in crate::engine) fn load_names(
        tensors: &ModelTensors,
        weight: &str,
        bias: Option<&str>,
        clipping_prefix: Option<&str>,
        stream: &Stream,
    ) -> Result<Self> {
        let weight = tensors.get(weight)?;
        let transposed_weight = weight.transpose(&[1, 0], stream)?;
        let bias = bias.map(|name| tensors.get_optional(name)).transpose()?.flatten();
        let clipping = clipping_prefix
            .map(|prefix| LinearClipping::load_optional(tensors, prefix))
            .transpose()?
            .flatten();
        Ok(Self { transposed_weight, bias, clipping })
    }

    pub fn forward(&self, input: &Array, stream: &Stream) -> Result<Array> {
        let clipped_input = self
            .clipping
            .as_ref()
            .map(|clipping| input.clip(&clipping.input_minimum, &clipping.input_maximum, stream))
            .transpose()?;
        let input = clipped_input.as_ref().unwrap_or(input);
        let output = input.matmul(&self.transposed_weight, stream)?;
        let output = match self.bias.as_ref() {
            Some(bias) => output.add(bias, stream),
            None => Ok(output),
        }?;
        match self.clipping.as_ref() {
            Some(clipping) => {
                output.clip(&clipping.output_minimum, &clipping.output_maximum, stream)
            },
            None => Ok(output),
        }
    }

    #[cfg(test)]
    pub(in crate::engine) fn from_arrays(
        weight: &Array,
        bias: Option<Array>,
        stream: &Stream,
    ) -> Result<Self> {
        Ok(Self {
            transposed_weight: weight.transpose(&[1, 0], stream)?,
            bias,
            clipping: None,
        })
    }

    #[cfg(test)]
    pub(in crate::engine) fn from_clipped_arrays(
        weight: &Array,
        input_bounds: (Array, Array),
        output_bounds: (Array, Array),
        stream: &Stream,
    ) -> Result<Self> {
        Ok(Self {
            transposed_weight: weight.transpose(&[1, 0], stream)?,
            bias: None,
            clipping: Some(LinearClipping {
                input_minimum: input_bounds.0,
                input_maximum: input_bounds.1,
                output_minimum: output_bounds.0,
                output_maximum: output_bounds.1,
            }),
        })
    }
}

impl LinearClipping {
    fn load_optional(tensors: &ModelTensors, prefix: &str) -> Result<Option<Self>> {
        let Some(input_minimum) = tensors.get_optional(&format!("{prefix}.input_min"))? else {
            return Ok(None);
        };
        Ok(Some(Self {
            input_minimum,
            input_maximum: tensors.get(&format!("{prefix}.input_max"))?,
            output_minimum: tensors.get(&format!("{prefix}.output_min"))?,
            output_maximum: tensors.get(&format!("{prefix}.output_max"))?,
        }))
    }
}