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
138
139
140

use super::{Kernels, template};
use crate::engine::{Error, Result};

impl Kernels {
    pub(crate) fn gated_delta_gates(
        &self,
        stream: &mirtal::Stream,
        inputs: [&mirtal::Array; 4],
    ) -> Result<[mirtal::Array; 2]> {
        let [alpha, beta, a_log, dt_bias] = inputs;
        let shape = alpha.shape()?;
        let dimensions = shape.dimensions().to_vec();
        if dimensions.len() != 3
            || beta.shape()? != shape
            || a_log.shape()? != dt_bias.shape()?
            || a_log.shape()?.dimensions() != [dimensions[2]]
        {
            return Err(Error::InvalidModel("Gated Delta gate shapes are incompatible".into()));
        }
        Ok(self.gated_delta_gates.dispatch(
            stream,
            inputs,
            &[
                mirtal::OutputSpec::new(shape.clone(), mirtal::DType::Float32),
                mirtal::OutputSpec::new(shape, mirtal::DType::Float32),
            ],
            &mirtal::Dispatch::new([alpha.len(), 1, 1], [256, 1, 1])
                .templates([template("HV", dimensions[2])?]),
        )?)
    }

    pub(crate) fn gated_delta_recurrence(
        &self,
        stream: &mirtal::Stream,
        inputs: [&mirtal::Array; 6],
    ) -> Result<[mirtal::Array; 2]> {
        let [query, key, value, .., state] = inputs;
        let key_shape = key.shape()?;
        let value_shape = value.shape()?;
        let key_dimensions = key_shape.dimensions().to_vec();
        let value_dimensions = value_shape.dimensions().to_vec();
        validate_recurrence(inputs, &key_dimensions, &value_dimensions)?;
        Ok(self.gated_delta_recurrence.dispatch(
            stream,
            inputs,
            &[
                mirtal::OutputSpec::new(value_shape, query.dtype()?),
                mirtal::OutputSpec::new(state.shape()?, state.dtype()?),
            ],
            &mirtal::Dispatch::new(
                [32, value_dimensions[3], value_dimensions[0] * value_dimensions[2]],
                [32, 4, 1],
            )
            .templates([
                mirtal::TemplateArg::dtype("InT", query.dtype()?),
                mirtal::TemplateArg::dtype("StT", state.dtype()?),
                template("DK", key_dimensions[3])?,
                template("DV", value_dimensions[3])?,
                template("HK", key_dimensions[2])?,
                template("HV", value_dimensions[2])?,
                template("STEPS", value_dimensions[1])?,
            ]),
        )?)
    }

    pub(crate) fn gated_delta_decode(
        &self,
        stream: &mirtal::Stream,
        inputs: [&mirtal::Array; 8],
        normalize: bool,
    ) -> Result<[mirtal::Array; 2]> {
        let [query, _, value, .., state] = inputs;
        let query_shape = query.shape()?;
        let value_shape = value.shape()?;
        let query_dimensions = query_shape.dimensions().to_vec();
        let value_dimensions = value_shape.dimensions().to_vec();
        validate_decode(inputs, &query_dimensions, &value_dimensions)?;
        Ok(self.gated_delta_decode.dispatch(
            stream,
            inputs,
            &[
                mirtal::OutputSpec::new(value_shape, query.dtype()?),
                mirtal::OutputSpec::new(state.shape()?, mirtal::DType::Float32),
            ],
            &mirtal::Dispatch::new(
                [256, 1, value_dimensions[0] * value_dimensions[2]],
                [256, 1, 1],
            )
            .templates([
                mirtal::TemplateArg::dtype("InT", query.dtype()?),
                mirtal::TemplateArg::dtype("StT", mirtal::DType::Float32),
                template("DK", query_dimensions[3])?,
                template("DV", value_dimensions[3])?,
                template("HK", query_dimensions[2])?,
                template("HV", value_dimensions[2])?,
                mirtal::TemplateArg::bool("NORMALIZE", normalize),
            ]),
        )?)
    }
}

fn validate_recurrence(inputs: [&mirtal::Array; 6], key: &[usize], value: &[usize]) -> Result<()> {
    let [query, keys, _, decay, update, state] = inputs;
    let ranks = key.len() == 4 && value.len() == 4;
    let expected_state = ranks.then(|| [value[0], value[2], value[3], key[3]]);
    let state_shape = state.shape()?;
    if query.shape()? != keys.shape()?
        || !ranks
        || key[0..2] != value[0..2]
        || !key[3].is_multiple_of(32)
        || decay.shape()? != update.shape()?
        || decay.shape()?.dimensions() != [value[0], value[1], value[2]]
        || expected_state.as_ref().is_none_or(|shape| state_shape.dimensions() != shape)
    {
        return Err(Error::InvalidModel("Gated Delta recurrence shapes are incompatible".into()));
    }
    Ok(())
}

fn validate_decode(inputs: [&mirtal::Array; 8], query: &[usize], value: &[usize]) -> Result<()> {
    let [queries, keys, _, alpha, beta, a_log, dt_bias, state] = inputs;
    let ranks = query.len() == 4 && value.len() == 4;
    let expected_state = ranks.then(|| [value[0], value[2], value[3], query[3]]);
    let state_shape = state.shape()?;
    if queries.shape()? != keys.shape()?
        || !ranks
        || query[1] != 1
        || value[1] != 1
        || !query[3].is_multiple_of(32)
        || alpha.shape()? != beta.shape()?
        || alpha.shape()?.dimensions() != [value[0], 1, value[2]]
        || a_log.shape()? != dt_bias.shape()?
        || a_log.shape()?.dimensions() != [value[2]]
        || state.dtype()? != mirtal::DType::Float32
        || expected_state.as_ref().is_none_or(|shape| state_shape.dimensions() != shape)
    {
        return Err(Error::InvalidModel("Gated Delta decode shapes are incompatible".into()));
    }
    Ok(())
}