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
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143

use models::{layout::SpatialMergeVisionConfig, vision::SpatialMergePreprocessedImage};

use super::{dimension, embedding::PatchEmbedding, layer::EncoderLayer, merger::PatchMerger};
use crate::engine::{Array, Error, ModelTensors, Result, Stream};

#[derive(Debug)]
pub struct SpatialMergeVisionTower {
    embedding: PatchEmbedding,
    layers: Vec<EncoderLayer>,
    merger: PatchMerger,
    patch_width: usize,
    output_hidden_size: usize,
    merge: usize,
}

impl SpatialMergeVisionTower {
    pub fn load(
        tensors: &ModelTensors,
        config: &SpatialMergeVisionConfig,
        stream: &Stream,
    ) -> Result<Self> {
        validate_config(config)?;
        let prefix = vision_prefix(tensors)?;
        let head_dim = config.hidden_size / config.num_attention_heads;
        let layers = (0..config.num_hidden_layers)
            .map(|layer| {
                EncoderLayer::load(
                    tensors,
                    &format!("{prefix}.blocks.{layer}"),
                    config.num_attention_heads,
                    head_dim,
                    stream,
                )
            })
            .collect::<Result<Vec<_>>>()?;
        Ok(Self {
            embedding: PatchEmbedding::load(tensors, config, prefix, stream)?,
            layers,
            merger: PatchMerger::load(
                tensors,
                config.hidden_size,
                config.spatial_merge_size,
                prefix,
                stream,
            )?,
            patch_width: config.in_channels
                * config.temporal_patch_size
                * config.patch_size
                * config.patch_size,
            output_hidden_size: config.output_hidden_size,
            merge: config.spatial_merge_size,
        })
    }

    pub fn forward_preprocessed(
        &self,
        image: &SpatialMergePreprocessedImage,
        stream: &Stream,
    ) -> Result<Array> {
        if image.grid_t != 1 {
            return Err(Error::InvalidModel(
                "spatial-merge vision Metal MVP accepts one still image, not video".into(),
            ));
        }
        let sequence =
            image.grid_height.checked_mul(image.grid_width).ok_or(Error::ShapeOverflow)?;
        if image.patches.len() != sequence * self.patch_width {
            return Err(Error::InvalidModel(
                "spatial-merge vision preprocessed patch buffer has an invalid length".into(),
            ));
        }
        let patches = Array::from_f32(
            &image.patches,
            &[
                1,
                dimension(sequence, "patch sequence")?,
                dimension(self.patch_width, "patch width")?,
            ],
        )?;
        let positions = position_ids(image, self.merge)?;
        let mut hidden = self.embedding.forward(&patches, image, stream)?;
        for layer in &self.layers {
            hidden = layer.forward(&hidden, &positions, stream)?;
        }
        let output = self.merger.forward(&hidden, stream)?;
        let expected = [
            1,
            dimension(image.soft_tokens, "soft tokens")?,
            dimension(self.output_hidden_size, "output hidden size")?,
        ];
        if output.shape()? != expected {
            return Err(Error::InvalidModel(format!(
                "spatial-merge vision tower output {:?} does not match {expected:?}",
                output.shape()?
            )));
        }
        Ok(output)
    }
}

fn vision_prefix(tensors: &ModelTensors) -> Result<&'static str> {
    for prefix in ["model.visual", "vision_tower"] {
        if tensors.contains(&format!("{prefix}.patch_embed.proj.weight"))? {
            return Ok(prefix);
        }
    }
    Err(Error::MissingTensor(
        "model.visual.patch_embed.proj.weight or vision_tower.patch_embed.proj.weight".into(),
    ))
}

fn position_ids(image: &SpatialMergePreprocessedImage, merge: usize) -> Result<Array> {
    let mut values = Vec::with_capacity(image.grid_height * image.grid_width * 2);
    for block_y in 0..image.grid_height / merge {
        for block_x in 0..image.grid_width / merge {
            for merge_y in 0..merge {
                for merge_x in 0..merge {
                    values.push(u32::try_from(block_y * merge + merge_y)?);
                    values.push(u32::try_from(block_x * merge + merge_x)?);
                }
            }
        }
    }
    Array::from_u32(
        &values,
        &[dimension(image.grid_height * image.grid_width, "position count")?, 2],
    )
}

fn validate_config(config: &SpatialMergeVisionConfig) -> Result<()> {
    if config.hidden_size == 0
        || config.num_attention_heads == 0
        || !config.hidden_size.is_multiple_of(config.num_attention_heads)
        || config.in_channels != 3
        || config.spatial_merge_size == 0
        || config.hidden_activation != "gelu_pytorch_tanh"
    {
        return Err(Error::InvalidModel(format!(
            "unsupported spatial-merge vision configuration: {config:?}"
        )));
    }
    Ok(())
}