libmir-metal 0.3.0

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
148
149
150

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

pub(super) fn batchable(contexts: &[&KvContext]) -> bool {
    let Some(first) = contexts.first().and_then(|context| context.paged.as_ref()) else {
        return false;
    };
    if first.key_scales.is_some() || first.value_scales.is_some() {
        return false;
    }
    let Ok(key_dtype) = first.key_pages.native().dtype() else {
        return false;
    };
    contexts.iter().skip(1).all(|context| {
        context.paged.as_ref().is_some_and(|paged| {
            paged.key_scales.is_none()
                && paged.value_scales.is_none()
                && paged.page_size == first.page_size
                && paged.context_tokens == first.context_tokens
                && paged.key_pages.native().dtype().is_ok_and(|dtype| dtype == key_dtype)
                && paged.value_pages.native().dtype().is_ok_and(|dtype| dtype == key_dtype)
        })
    })
}

pub(super) fn rows(
    queries: &[&Array],
    contexts: &[&KvContext],
    scale: f32,
    stream: &Stream,
) -> Result<Vec<Array>> {
    queries
        .iter()
        .zip(contexts)
        .map(|(query, context)| {
            let paged = context.paged.as_ref().ok_or_else(|| {
                super::super::Error::InvalidModel("paged context is missing".into())
            })?;
            query.paged_scaled_dot_product_attention_with_scratch(
                paged.attention(),
                paged.scratch(),
                scale,
                stream,
            )
        })
        .collect()
}

pub(super) fn batched(
    queries: &[&Array],
    contexts: &[&KvContext],
    scale: f32,
    stream: &Stream,
) -> Result<Vec<Array>> {
    if !batchable(contexts) {
        return Err(super::super::Error::InvalidModel(
            "paged contexts are not batch compatible".into(),
        ));
    }
    let mut chunks = Vec::new();
    for (queries, contexts) in queries
        .chunks(super::super::kernels::BATCHED_PAGED_ROWS)
        .zip(contexts.chunks(super::super::kernels::BATCHED_PAGED_ROWS))
    {
        chunks.push(chunk(queries, contexts, scale, stream)?);
    }
    let output = if chunks.len() == 1 {
        chunks.pop().ok_or(super::super::Error::ShapeOverflow)?
    } else {
        let refs = chunks.iter().collect::<Vec<_>>();
        Array::concatenate(&refs, 0, stream)?
    };
    let shape = output.shape()?;
    (0..shape[0])
        .map(|row| {
            output.slice(
                &[usize::try_from(row)?, 0, 0, 0],
                &[
                    usize::try_from(row + 1)?,
                    usize::try_from(shape[1])?,
                    usize::try_from(shape[2])?,
                    usize::try_from(shape[3])?,
                ],
                stream,
            )
        })
        .collect()
}

fn chunk(
    queries: &[&Array],
    contexts: &[&KvContext],
    scale: f32,
    stream: &Stream,
) -> Result<Array> {
    let first = paged(contexts[0])?;
    let queries = Array::concatenate(queries, 0, stream)?;
    let pages = first.context_tokens.div_ceil(first.page_size);
    let mut tables = Vec::with_capacity(contexts.len());
    let mut dependencies = Vec::with_capacity(contexts.len());
    let mut capacities = Vec::with_capacity(contexts.len());
    let mut keys = [first.key_pages.native(); 8];
    let mut values = [first.value_pages.native(); 8];
    for (row, context) in contexts.iter().enumerate() {
        let context = paged(context)?;
        tables.push(context.page_table.slice(&[0], &[pages], stream)?);
        dependencies.push(&context.page_dependency);
        capacities.push(u32::try_from(context.key_pages.native().shape()?.dimensions()[1])?);
        keys[row] = context.key_pages.native();
        values[row] = context.value_pages.native();
    }
    let table_refs = tables.iter().collect::<Vec<_>>();
    let tables = Array::concatenate(&table_refs, 0, stream)?;
    let dependencies = Array::concatenate(&dependencies, 0, stream)?;
    let capacities = Array::from_u32(&capacities, &[i32::try_from(capacities.len())?])?;
    let output = stream.batched_paged_attention(
        [
            queries.native(),
            keys[0],
            keys[1],
            keys[2],
            keys[3],
            keys[4],
            keys[5],
            keys[6],
            keys[7],
            values[0],
            values[1],
            values[2],
            values[3],
            values[4],
            values[5],
            values[6],
            values[7],
            tables.native(),
            dependencies.native(),
            capacities.native(),
        ],
        first.page_size,
        first.context_tokens,
        scale,
    )?;
    Array::from_native(output)
}

fn paged(context: &KvContext) -> Result<&super::super::PagedKvContext> {
    context
        .paged
        .as_ref()
        .ok_or_else(|| super::super::Error::InvalidModel("paged context is missing".into()))
}