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::{HybridMoeLayerConfig, weights::AttentionWeights};
use crate::engine::{
    Array, FusedAttention, FusedKeyValue, KvCache, PagedContextMode, Result, RopeOptions, Stream,
    native_paged_attention_mode, paged_attention_min_context,
};

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

pub(super) fn forward_decode(
    input: &Array,
    weights: &AttentionWeights,
    config: HybridMoeLayerConfig,
    fused_attention: Option<&FusedAttention>,
    fused_key_value: Option<&FusedKeyValue>,
    context: DecodeContext<'_>,
) -> Result<Array> {
    let DecodeContext { cache, position, causal, stream } = context;
    let sequence = input.shape()?.get(1).copied().ok_or_else(|| {
        crate::engine::Error::InvalidModel("attention input has no sequence axis".into())
    })?;
    let fused_attention = (sequence == 1)
        .then_some(fused_attention)
        .flatten()
        .map(|fused| fused.forward(input, stream))
        .transpose()?;
    let (queries, raw_keys, raw_values) = if let Some(output) = fused_attention {
        (output.query, output.key, output.value)
    } else {
        let fused_key_value = (sequence == 1)
            .then_some(fused_key_value)
            .flatten()
            .map(|fused| fused.forward(input, stream))
            .transpose()?;
        let (raw_keys, raw_values) = match fused_key_value {
            Some((key, value)) => (key, Some(value)),
            None => (
                weights.key.forward(input, stream)?,
                weights.value.as_ref().map(|value| value.forward(input, stream)).transpose()?,
            ),
        };
        (weights.query.forward(input, stream)?, raw_keys, raw_values)
    };
    let queries =
        queries.reshape(&[1, sequence, config.attention_heads, config.head_dim], stream)?;
    let queries = weights.query_norm.apply(&queries, config.rms_norm_eps, stream)?;
    let queries =
        rope_layout(&queries, weights.rope_frequencies.as_ref(), config, position, stream)?;

    let raw_keys = raw_keys.reshape(&[1, sequence, config.kv_heads, config.head_dim], stream)?;
    let values = if config.use_k_eq_v {
        raw_keys
            .rms_norm_unit(config.rms_norm_eps, stream)?
            .transpose(&[0, 2, 1, 3], stream)?
    } else {
        let raw_values = raw_values.ok_or_else(|| {
            crate::engine::Error::InvalidModel("missing hybrid MoE value projection".into())
        })?;
        raw_values
            .reshape(&[1, sequence, config.kv_heads, config.head_dim], stream)?
            .rms_norm_unit(config.rms_norm_eps, stream)?
            .transpose(&[0, 2, 1, 3], stream)?
    };
    let keys = weights.key_norm.apply(&raw_keys, config.rms_norm_eps, stream)?;
    let keys = rope_layout(&keys, weights.rope_frequencies.as_ref(), config, position, stream)?;
    let output = match cache {
        Some(cache) => {
            let mode = if sequence == 1 {
                native_paged_attention_mode(
                    config.head_dim,
                    config.attention_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,
            )?;
            if let Some(paged) = context.paged {
                queries.paged_scaled_dot_product_attention_with_scratch(
                    paged.attention(),
                    paged.scratch(),
                    1.0,
                    stream,
                )?
            } else if let Some(mask) = context.mask.as_ref() {
                queries.masked_scaled_dot_product_attention(
                    &context.keys, &context.values, 1.0, mask, stream,
                )?
            } else {
                queries.scaled_dot_product_attention(
                    &context.keys, &context.values, 1.0, causal, stream,
                )?
            }
        },
        None => queries.scaled_dot_product_attention(&keys, &values, 1.0, causal, stream)?,
    };
    let output = output.transpose(&[0, 2, 1, 3], stream)?;
    let output_width = config.attention_heads * config.head_dim;
    let output = output.reshape(&[1, sequence, output_width], stream)?;
    weights.output.forward(&output, stream)
}

pub(super) fn rope_layout(
    input: &Array,
    frequencies: Option<&Array>,
    config: HybridMoeLayerConfig,
    position: i32,
    stream: &Stream,
) -> Result<Array> {
    let input = input.transpose(&[0, 2, 1, 3], stream)?;
    frequencies.map_or_else(
        || {
            input.rope(
                RopeOptions {
                    dimensions: config.rope_dimensions,
                    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)
        },
    )
}