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

use super::{config::DenseSwiGluLayerConfig, weights::AttentionWeights};
use crate::engine::{
    Array, FusedAttention, KvCache, NormWeight, PagedContextMode, Result, RopeOptions, Stream,
    native_paged_attention_mode, paged_attention_min_context,
};

pub(super) struct AttentionContext<'a> {
    pub cache: &'a mut KvCache,
    pub position: i32,
    pub causal: bool,
    pub stream: &'a Stream,
}

pub(super) fn forward(
    input: &Array,
    weights: &AttentionWeights,
    fused_attention: Option<&FusedAttention>,
    config: DenseSwiGluLayerConfig,
    context: AttentionContext<'_>,
) -> Result<Array> {
    let AttentionContext { cache, position, causal, stream } = context;
    let sequence = input.shape()?.get(1).copied().ok_or_else(|| {
        crate::engine::Error::InvalidModel(
            "dense SwiGLU attention input has no sequence axis".into(),
        )
    })?;
    let fused = (sequence == 1)
        .then_some(fused_attention)
        .flatten()
        .map(|fused| fused.forward(input, stream))
        .transpose()?;
    let (queries, keys, values) = match fused {
        Some(output) => (
            output.query,
            output.key,
            output.value.ok_or_else(|| {
                crate::engine::Error::InvalidModel("fused attention omitted values".into())
            })?,
        ),
        None => (
            weights.query.forward(input, stream)?,
            weights.key.forward(input, stream)?,
            weights.value.forward(input, stream)?,
        ),
    };
    let queries = queries.reshape(&[1, sequence, config.heads, config.head_dim], stream)?;
    let queries = normalize(queries, weights.query_norm.as_ref(), config.rms_norm_eps, stream)?;
    let queries =
        rope_layout(&queries, weights.rope_frequencies.as_ref(), config, position, stream)?;
    let keys = keys.reshape(&[1, sequence, config.kv_heads, config.head_dim], stream)?;
    let keys = normalize(keys, weights.key_norm.as_ref(), config.rms_norm_eps, stream)?;
    let keys = rope_layout(&keys, weights.rope_frequencies.as_ref(), config, position, stream)?;
    let values = values
        .reshape(&[1, sequence, config.kv_heads, config.head_dim], stream)?
        .transpose(&[0, 2, 1, 3], stream)?;
    let mode = if sequence == 1 {
        native_paged_attention_mode(
            config.head_dim,
            config.heads,
            config.kv_heads,
            usize::try_from(position)? + 1,
            stream.config().cache.force_native_paged_attention,
        )
    } else {
        PagedContextMode::View
    };
    let context = cache.update_for_attention_mode(
        &keys,
        &values,
        stream,
        paged_attention_min_context(stream),
        mode,
    )?;
    let output = match context.paged {
        Some(paged) => queries.paged_scaled_dot_product_attention_with_scratch(
            paged.attention(),
            paged.scratch(),
            config.attention_scale,
            stream,
        )?,
        None => queries.scaled_dot_product_attention(
            &context.keys,
            &context.values,
            config.attention_scale,
            causal,
            stream,
        )?,
    };
    let output = output.transpose(&[0, 2, 1, 3], stream)?;
    let output_width = config.heads * config.head_dim;
    weights
        .output
        .forward(&output.reshape(&[1, sequence, output_width], stream)?, stream)
}

pub(super) fn normalize(
    input: Array,
    weight: Option<&NormWeight>,
    eps: f32,
    stream: &Stream,
) -> Result<Array> {
    match weight {
        Some(weight) => weight.apply(&input, eps, stream),
        None => Ok(input),
    }
}

pub(super) fn rope_layout(
    input: &Array,
    frequencies: Option<&Array>,
    config: DenseSwiGluLayerConfig,
    position: i32,
    stream: &Stream,
) -> Result<Array> {
    let input = input.transpose(&[0, 2, 1, 3], stream)?;
    frequencies.map_or_else(
        || {
            input.rope(
                RopeOptions {
                    dimensions: config.head_dim,
                    traditional: false,
                    base: Some(config.rope_base),
                    scale: 1.0,
                    offset: position,
                },
                stream,
            )
        },
        |frequencies| {
            input.rope_with_frequencies(config.head_dim, false, frequencies, position, stream)
        },
    )
}