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

use models::{layout::PooledVisionConfig, vision::PooledPreprocessedImage};

use super::{embedding::PatchEmbedding, layer::EncoderLayer, pooler::VisionPooler};
use crate::engine::{Array, Error, ModelTensors, Result, Stream};

#[derive(Debug)]
pub struct PooledVisionTower {
    embedding: PatchEmbedding,
    layers: Vec<EncoderLayer>,
    pooler: VisionPooler,
    patch_width: usize,
    pub(super) bidirectional_image_attention: bool,
}

impl PooledVisionTower {
    pub fn load(
        tensors: &ModelTensors,
        config: &PooledVisionConfig,
        stream: &Stream,
    ) -> Result<Self> {
        validate_config(config)?;
        let eps = config.rms_norm_eps.to_string().parse::<f32>()?;
        let layers = (0..config.num_hidden_layers)
            .map(|layer| {
                EncoderLayer::load(
                    tensors,
                    &format!("model.vision_tower.encoder.layers.{layer}"),
                    config.num_attention_heads,
                    config.num_key_value_heads,
                    config.head_dim,
                    config.rope_theta,
                    eps,
                    stream,
                )
            })
            .collect::<Result<Vec<_>>>()?;
        Ok(Self {
            embedding: PatchEmbedding::load(tensors, stream)?,
            layers,
            pooler: VisionPooler::load(
                tensors,
                config.hidden_size,
                config.pooling_kernel_size,
                config.standardize,
                eps,
                stream,
            )?,
            patch_width: 3 * config.patch_size * config.patch_size,
            bidirectional_image_attention: config.bidirectional_image_attention,
        })
    }

    /// Encodes already extracted RGB patches with explicit `(row, column)`
    /// position IDs.
    ///
    /// Padding must be removed before this call. Every image in the batch must
    /// use the same `grid_height` and `grid_width`; the returned tensor
    /// stays accelerator-resident.
    pub fn forward(
        &self,
        patches: &Array,
        position_ids: &Array,
        grid_height: usize,
        grid_width: usize,
        stream: &Stream,
    ) -> Result<Array> {
        let shape = patches.shape()?;
        if shape.len() != 3 || usize::try_from(shape[2])? != self.patch_width {
            return Err(Error::InvalidModel(format!(
                "pooled vision expected RGB patch width {}, got {shape:?}",
                self.patch_width
            )));
        }
        let mut hidden = self.embedding.forward(patches, position_ids, stream)?;
        for layer in &self.layers {
            hidden = layer.forward(&hidden, position_ids, stream)?;
        }
        self.pooler.forward(&hidden, grid_height, grid_width, stream)
    }

    /// Uploads one preprocessed image and encodes it on the Metal stream.
    pub fn forward_preprocessed(
        &self,
        image: &PooledPreprocessedImage,
        stream: &Stream,
    ) -> Result<Array> {
        let patches =
            image.grid_height.checked_mul(image.grid_width).ok_or(Error::ShapeOverflow)?;
        let patch_values = patches.checked_mul(self.patch_width).ok_or(Error::ShapeOverflow)?;
        if image.patches.len() != patch_values || image.position_ids.len() != patches * 2 {
            return Err(Error::InvalidModel(format!(
                "pooled vision preprocessed image has inconsistent buffers for grid {}x{}",
                image.grid_height, image.grid_width
            )));
        }
        let patches = i32::try_from(patches)?;
        let patch_width = i32::try_from(self.patch_width)?;
        let patch_array = Array::from_f32(&image.patches, &[1, patches, patch_width])?;
        let positions = Array::from_u32(&image.position_ids, &[1, patches, 2])?;
        let output =
            self.forward(&patch_array, &positions, image.grid_height, image.grid_width, stream)?;
        if usize::try_from(output.shape()?[1])? != image.soft_tokens {
            return Err(Error::InvalidModel("pooled vision soft-token count mismatch".into()));
        }
        Ok(output)
    }
}

fn validate_config(config: &PooledVisionConfig) -> Result<()> {
    let projected = config
        .num_attention_heads
        .checked_mul(config.head_dim)
        .ok_or(Error::ShapeOverflow)?;
    if config.hidden_size != projected
        || config.num_key_value_heads == 0
        || !config.num_attention_heads.is_multiple_of(config.num_key_value_heads)
        || config.pooling_kernel_size == 0
        || config.hidden_activation != "gelu_pytorch_tanh"
    {
        return Err(Error::InvalidModel(format!(
            "unsupported pooled vision dimensions or activation: hidden={}, heads={}, kv_heads={}, head_dim={}, pool={}, activation={}",
            config.hidden_size,
            config.num_attention_heads,
            config.num_key_value_heads,
            config.head_dim,
            config.pooling_kernel_size,
            config.hidden_activation
        )));
    }
    Ok(())
}