libmir-metal 0.1.0

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

use mirtal::{DType, Graph, Shape};

use super::super::{Array, Error, Result, Stream};

#[derive(Debug, Clone, Copy)]
pub struct RopeOptions {
    pub dimensions: i32,
    pub traditional: bool,
    pub base: Option<f32>,
    pub scale: f32,
    pub offset: i32,
}

impl Array {
    pub(crate) fn proportional_rope_frequencies(
        head_dim: i32,
        rope_dimensions: i32,
        base: f32,
        stream: &Stream,
    ) -> Result<Self> {
        dimensions(head_dim)?;
        dimensions(rope_dimensions)?;
        if rope_dimensions > head_dim {
            return Err(invalid("RoPE dimensions exceed head dimensions"));
        }
        positive(base, "base")?;
        let graph = stream.native().graph();
        let exponents = graph.arange(0.0, as_f32(rope_dimensions)?, 2.0, DType::Float32)?;
        let exponents = graph.divide(&exponents, &scalar(graph, as_f32(head_dim)?)?)?;
        let frequencies = graph.power(&scalar(graph, base)?, &exponents)?;
        let remaining = usize::try_from(head_dim / 2 - rope_dimensions / 2)?;
        if remaining == 0 {
            return Self::from_native(frequencies);
        }
        let tail = graph.full(&Shape::new([remaining])?, f32::INFINITY, DType::Float32)?;
        Self::from_native(graph.concatenate(&[&frequencies, &tail], 0)?)
    }

    pub(crate) fn piecewise_rope_frequencies(
        rotary_dimensions: i32,
        base: f32,
        factor: f32,
        low_frequency_factor: f32,
        high_frequency_factor: f32,
        original_context_len: i32,
        stream: &Stream,
    ) -> Result<Self> {
        dimensions(rotary_dimensions)?;
        for (value, name) in [
            (base, "base"),
            (factor, "factor"),
            (low_frequency_factor, "low frequency factor"),
            (high_frequency_factor, "high frequency factor"),
        ] {
            positive(value, name)?;
        }
        if high_frequency_factor <= low_frequency_factor || original_context_len <= 0 {
            return Err(invalid("invalid piecewise RoPE interpolation range"));
        }
        let graph = stream.native().graph();
        let exponents = graph.arange(0.0, as_f32(rotary_dimensions)?, 2.0, DType::Float32)?;
        let exponents = graph.divide(&exponents, &scalar(graph, as_f32(rotary_dimensions)?)?)?;
        let frequencies = graph.power(&scalar(graph, base)?, &exponents)?;
        let inverse = graph.reciprocal(&frequencies)?;
        let wavelength = graph.multiply(&scalar(graph, std::f32::consts::TAU)?, &frequencies)?;
        let ratio = graph.divide(&scalar(graph, as_f32(original_context_len)?)?, &wavelength)?;
        let smooth = graph.subtract(&ratio, &scalar(graph, low_frequency_factor)?)?;
        let smooth =
            graph.divide(&smooth, &scalar(graph, high_frequency_factor - low_frequency_factor)?)?;
        let smooth = graph.maximum(&smooth, &scalar(graph, 0.0)?)?;
        let smooth = graph.minimum(&smooth, &scalar(graph, 1.0)?)?;
        let slowed = graph.divide(&inverse, &scalar(graph, factor)?)?;
        let inverse_smooth = graph.subtract(&scalar(graph, 1.0)?, &smooth)?;
        let adjusted = graph
            .add(&graph.multiply(&slowed, &inverse_smooth)?, &graph.multiply(&inverse, &smooth)?)?;
        Self::from_native(graph.reciprocal(&adjusted)?)
    }

    pub fn rope(&self, options: RopeOptions, stream: &Stream) -> Result<Self> {
        Self::from_native(stream.native().graph().rope(
            self.native(),
            mirtal::RopeOptions {
                dimensions: usize::try_from(options.dimensions)?,
                traditional: options.traditional,
                base: options.base,
                scale: options.scale,
                offset: usize::try_from(options.offset)?,
            },
        )?)
    }

    pub fn rope_with_frequencies(
        &self,
        dimensions: i32,
        traditional: bool,
        frequencies: &Self,
        offset: i32,
        stream: &Stream,
    ) -> Result<Self> {
        Self::from_native(stream.native().graph().rope_with_frequencies(
            self.native(),
            frequencies.native(),
            mirtal::FrequencyRopeOptions {
                dimensions: usize::try_from(dimensions)?,
                traditional,
                offset: usize::try_from(offset)?,
            },
        )?)?
        .astype_like(self, stream)
    }
}

fn scalar(graph: Graph<'_>, value: f32) -> Result<mirtal::Array> {
    Ok(graph.full(&Shape::new([])?, value, DType::Float32)?)
}

fn dimensions(value: i32) -> Result<()> {
    if value <= 0 || value % 2 != 0 {
        return Err(invalid("RoPE dimensions must be positive and even"));
    }
    Ok(())
}

fn positive(value: f32, name: &str) -> Result<()> {
    if !value.is_finite() || value <= 0.0 {
        return Err(invalid(format!("{name} must be finite and positive")));
    }
    Ok(())
}

fn as_f32(value: i32) -> Result<f32> {
    Ok(f32::from(u16::try_from(value)?))
}

fn invalid(message: impl Into<String>) -> Error {
    Error::InvalidModel(message.into())
}