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
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

use std::time::Instant;

use super::HybridMoeModel;
use crate::engine::{
    Array, DecoderCache, Error, ImageTokenSpan, Result, Stream,
    decoder::{LayerContext, LayerLoopOptions, forward_layers},
};

impl HybridMoeModel {
    pub fn forward_prefill(
        &self,
        token_ids: &Array,
        cache: &mut DecoderCache,
        position: i32,
        stream: &Stream,
    ) -> Result<Array> {
        self.forward_hidden(token_ids, cache, position, true, stream)
    }

    /// Runs an initial multimodal prefill without copying image embeddings
    /// back to the host.
    pub fn forward_multimodal_prefill(
        &self,
        token_ids: &Array,
        image_embeddings: &Array,
        image: ImageTokenSpan,
        bidirectional_image_attention: bool,
        cache: &mut DecoderCache,
        stream: &Stream,
    ) -> Result<Array> {
        if cache.cached_tokens()? != 0 {
            return Err(Error::InvalidModel(
                "multimodal prefill requires an empty decoder cache".into(),
            ));
        }
        let hidden = self.mixed_embeddings(token_ids, image_embeddings, image, stream)?;
        let image = bidirectional_image_attention.then_some(image);
        self.forward_embedded(hidden, cache, 0, true, image, stream)
    }

    pub fn mixed_embeddings(
        &self,
        token_ids: &Array,
        image_embeddings: &Array,
        image: ImageTokenSpan,
        stream: &Stream,
    ) -> Result<Array> {
        let token_shape = token_ids.shape()?;
        if token_shape.len() != 2 || token_shape[0] != 1 {
            return Err(Error::InvalidModel(format!(
                "multimodal token IDs must have shape [1, sequence], got {token_shape:?}"
            )));
        }
        let sequence = usize::try_from(token_shape[1])?;
        let image = ImageTokenSpan::new(image.start..image.end, sequence)?;
        let image_shape = image_embeddings.shape()?;
        let expected = [1, i32::try_from(image.len())?, i32::try_from(self.hidden_size)?];
        if image_shape != expected {
            return Err(Error::InvalidModel(format!(
                "image embeddings {image_shape:?} do not match expected {expected:?}"
            )));
        }
        let hidden = self
            .embedding
            .lookup(token_ids, stream)?
            .multiply_scalar(self.embed_scale, stream)?;
        hidden.slice_update(
            image_embeddings,
            &[0, image.start, 0],
            &[1, image.end, self.hidden_size],
            stream,
        )
    }

    pub(super) fn forward_hidden(
        &self,
        token_ids: &Array,
        cache: &mut DecoderCache,
        position: i32,
        causal: bool,
        stream: &Stream,
    ) -> Result<Array> {
        let profile = !causal && super::profile_components(stream);
        let started = Instant::now();
        let hidden = self
            .embedding
            .lookup(token_ids, stream)?
            .multiply_scalar(self.embed_scale, stream)?;
        if profile {
            hidden.async_eval()?;
            stream.synchronize()?;
            tracing::debug!(
                component = "embedding",
                milliseconds = started.elapsed().as_secs_f64() * 1_000.0,
                "MLX hybrid MoE component profile"
            );
        }
        self.forward_embedded(hidden, cache, position, causal, None, stream)
    }

    fn forward_embedded(
        &self,
        hidden: Array,
        cache: &mut DecoderCache,
        position: i32,
        causal: bool,
        image: Option<ImageTokenSpan>,
        stream: &Stream,
    ) -> Result<Array> {
        let profile = stream.config().diagnostics.profile_layers;
        let evaluation_step = causal
            .then_some(stream.config().diagnostics.prefill_evaluation_layers)
            .flatten();
        forward_layers(
            &self.layers,
            hidden,
            cache,
            LayerContext {
                position,
                causal,
                positions: None,
                image,
                stream,
            },
            LayerLoopOptions::new(profile, evaluation_step, false),
        )
    }
}