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
151
152
153
154

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

#[derive(Clone, Copy, Debug, Eq, PartialEq)]
struct RowKey {
    sequence: usize,
    context: usize,
    query_heads: usize,
    kv_heads: usize,
    head_dim: usize,
    dtype: u8,
    causal: bool,
    fragmented: bool,
}

#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq, serde::Deserialize, serde::Serialize)]
pub struct BatchAttentionKey {
    pub batch: usize,
    pub sequence: usize,
    pub context_bucket: usize,
    pub query_heads: usize,
    pub kv_heads: usize,
    pub head_dim: usize,
    pub dtype: u8,
    pub causal: bool,
    pub fragmented: bool,
}

pub(super) fn key(
    queries: &[&Array],
    contexts: &[&KvContext],
    causal: bool,
) -> Result<Option<BatchAttentionKey>> {
    if queries.len() < 2 || queries.len() != contexts.len() {
        return Ok(None);
    }
    let Some(first) = row_key(queries[0], contexts[0], causal)? else {
        return Ok(None);
    };
    for (&query, &context) in queries[1..].iter().zip(&contexts[1..]) {
        if row_key(query, context, causal)? != Some(first) {
            return Ok(None);
        }
    }
    Ok(Some(BatchAttentionKey {
        batch: queries.len(),
        sequence: first.sequence,
        context_bucket: context_bucket(first.context),
        query_heads: first.query_heads,
        kv_heads: first.kv_heads,
        head_dim: first.head_dim,
        dtype: first.dtype,
        causal,
        fragmented: first.fragmented,
    }))
}

pub(in crate::engine) fn compatible_groups(
    queries: &[&Array],
    contexts: &[&KvContext],
    causal: bool,
) -> Result<Vec<Vec<usize>>> {
    if queries.len() != contexts.len() {
        return Ok(Vec::new());
    }
    let mut groups = Vec::<(RowKey, Vec<usize>)>::new();
    let mut singles = Vec::new();
    for (index, (&query, &context)) in queries.iter().zip(contexts).enumerate() {
        let Some(key) = row_key(query, context, causal)? else {
            singles.push(vec![index]);
            continue;
        };
        if let Some((_, rows)) = groups.iter_mut().find(|(candidate, _)| *candidate == key) {
            rows.push(index);
        } else {
            groups.push((key, vec![index]));
        }
    }
    Ok(groups.into_iter().map(|(_, rows)| rows).chain(singles).collect())
}

pub(super) fn fallback(key: BatchAttentionKey, paged: bool) -> super::BatchAttentionExecution {
    if paged
        && key.fragmented
        && key.sequence == 1
        && key.context_bucket >= 1_024
        && key.head_dim <= 256
        && key.head_dim.is_multiple_of(32)
        && key.kv_heads > 0
        && key.query_heads.is_multiple_of(key.kv_heads)
        && key.query_heads / key.kv_heads <= 32
    {
        super::BatchAttentionExecution::PagedRows
    } else {
        super::BatchAttentionExecution::Rows
    }
}

pub(super) const fn prefer_paged_batched(key: BatchAttentionKey, batchable: bool) -> bool {
    batchable && key.fragmented && key.sequence == 1 && key.context_bucket >= 8_192
}

fn row_key(query: &Array, context: &KvContext, causal: bool) -> Result<Option<RowKey>> {
    if context
        .paged
        .as_ref()
        .is_some_and(|paged| paged.key_scales.is_some() || paged.value_scales.is_some())
    {
        return Ok(None);
    }
    let dtype = dtype_key(query.native().dtype()?);
    let query = query.native().shape()?;
    let keys = context.keys.native().shape()?;
    let values = context.values.native().shape()?;
    let query = query.dimensions();
    let keys = keys.dimensions();
    let values = values.dimensions();
    let compatible = query.len() == 4
        && keys.len() == 4
        && values == keys
        && query[0] == 1
        && query[2] == 1
        && keys[0] == 1
        && query[2] > 0
        && keys[2] > 0
        && keys[1] > 0
        && query[1].is_multiple_of(keys[1])
        && query[3] == keys[3];
    Ok(compatible.then(|| RowKey {
        sequence: query[2],
        context: keys[2],
        query_heads: query[1],
        kv_heads: keys[1],
        head_dim: query[3],
        dtype,
        causal,
        fragmented: context.paged.as_ref().is_some_and(|paged| paged.fragmented),
    }))
}

fn context_bucket(tokens: usize) -> usize {
    tokens.max(1_024).checked_next_power_of_two().unwrap_or(usize::MAX)
}

const fn dtype_key(dtype: mirtal::DType) -> u8 {
    match dtype {
        mirtal::DType::Bool => 0,
        mirtal::DType::Uint8 => 1,
        mirtal::DType::Uint32 => 2,
        mirtal::DType::Int32 => 3,
        mirtal::DType::Float16 => 4,
        mirtal::DType::Bfloat16 => 5,
        mirtal::DType::Float32 => 6,
    }
}