libmir-metal 0.2.0

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

use crate::engine::{Array, NormWeight, QuantizedLinear};

mod load;

#[derive(Debug)]
pub(super) struct AttentionWeights {
    pub(super) query: QuantizedLinear,
    pub(super) key: QuantizedLinear,
    pub(super) value: Option<QuantizedLinear>,
    pub(super) output: QuantizedLinear,
    pub(super) query_norm: NormWeight,
    pub(super) key_norm: NormWeight,
    pub(super) rope_frequencies: Option<Array>,
}

#[derive(Debug)]
pub(super) struct DenseWeights {
    pub(super) gate: QuantizedLinear,
    pub(super) up: QuantizedLinear,
    pub(super) down: QuantizedLinear,
}

#[derive(Debug)]
pub(super) struct RouterWeights {
    pub(super) projection: QuantizedLinear,
    pub(super) norm_scale: Array,
    pub(super) expert_scale: Array,
}

#[derive(Debug)]
pub(super) struct ExpertWeights {
    pub(super) gate: QuantizedLinear,
    pub(super) up: QuantizedLinear,
    pub(super) down: QuantizedLinear,
}

#[derive(Debug)]
pub(super) struct LayerWeights {
    pub(super) input_norm: NormWeight,
    pub(super) post_attention_norm: NormWeight,
    pub(super) pre_dense_norm: NormWeight,
    pub(super) post_dense_norm: NormWeight,
    pub(super) pre_expert_norm: NormWeight,
    pub(super) post_expert_norm: NormWeight,
    pub(super) post_feed_forward_norm: NormWeight,
    pub(super) layer_scalar: Array,
    pub(super) attention: AttentionWeights,
    pub(super) dense: DenseWeights,
    pub(super) router: RouterWeights,
    pub(super) experts: ExpertWeights,
}