libmir-metal 0.3.0

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::weights::{BlockProjectionLayout, BlockQuantization, TensorBinding, TensorStorage};

use super::{ClampedRoutedConfig, ExpertLayout};
use crate::engine::{Array, Dtype, Error, ModelTensors, Result};

pub(super) fn native_layout(
    tensors: &ModelTensors,
    gate_up: &TensorBinding,
    down: &TensorBinding,
    config: ClampedRoutedConfig,
) -> Result<ExpertLayout> {
    let Some(BlockProjectionLayout::FusedGateUpBank { experts, interleaved: true }) =
        gate_up.block_projection_layout()
    else {
        return Err(invalid(gate_up, "requires an interleaved gate/up expert bank"));
    };
    let (gate_up_scales, gate_up_bias) =
        companions(gate_up, BlockProjectionLayout::FusedGateUpBank { experts, interleaved: true })?;
    let (down_scales, down_bias) =
        companions(down, BlockProjectionLayout::MatrixBank { matrices: experts })?;
    let hidden = usize::try_from(config.hidden)?;
    let intermediate = usize::try_from(config.intermediate)?;
    let gate_rows = intermediate.checked_mul(2).ok_or(Error::ShapeOverflow)?;
    let layout = ExpertLayout::Native {
        gate_up_blocks: tensors.get(&gate_up.source)?,
        gate_up_scales: tensors.get(gate_up_scales)?,
        gate_up_bias: tensors.get(gate_up_bias)?,
        down_blocks: tensors.get(&down.source)?,
        down_scales: tensors.get(down_scales)?,
        down_bias: tensors.get(down_bias)?,
    };
    validate(&layout, experts, hidden, intermediate, gate_rows, gate_up, down)?;
    Ok(layout)
}

fn companions(binding: &TensorBinding, expected: BlockProjectionLayout) -> Result<(&str, &str)> {
    let TensorStorage::BlockQuantized {
        format: BlockQuantization::MXFP4,
        scales,
        bias: Some(bias),
        ..
    } = &binding.storage
    else {
        return Err(invalid(binding, "requires MXFP4 scales and BF16 bias"));
    };
    if binding.block_projection_layout() != Some(expected) {
        return Err(invalid(binding, "matrix-bank layout differs from routed role"));
    }
    Ok((scales, bias))
}

#[allow(clippy::too_many_arguments)]
fn validate(
    layout: &ExpertLayout,
    experts: usize,
    hidden: usize,
    intermediate: usize,
    gate_rows: usize,
    gate: &TensorBinding,
    down: &TensorBinding,
) -> Result<()> {
    let ExpertLayout::Native {
        gate_up_blocks,
        gate_up_scales,
        gate_up_bias,
        down_blocks,
        down_scales,
        down_bias,
    } = layout
    else {
        return Err(Error::InvalidModel("expected native MXFP4 experts".into()));
    };
    require(gate_up_blocks, Dtype::Uint8, &[experts, gate_rows, hidden / 32, 16], gate)?;
    require(gate_up_scales, Dtype::Uint8, &[experts, gate_rows, hidden / 32], gate)?;
    require(gate_up_bias, Dtype::Bfloat16, &[experts, gate_rows], gate)?;
    require(down_blocks, Dtype::Uint8, &[experts, hidden, intermediate / 32, 16], down)?;
    require(down_scales, Dtype::Uint8, &[experts, hidden, intermediate / 32], down)?;
    require(down_bias, Dtype::Bfloat16, &[experts, hidden], down)
}

fn require(array: &Array, dtype: Dtype, shape: &[usize], binding: &TensorBinding) -> Result<()> {
    let expected = shape
        .iter()
        .copied()
        .map(i32::try_from)
        .collect::<std::result::Result<Vec<_>, _>>()?;
    if array.dtype()? == dtype && array.shape()? == expected {
        Ok(())
    } else {
        Err(invalid(binding, "MXFP4 expert companion dtype or shape differs"))
    }
}

fn invalid(binding: &TensorBinding, reason: &str) -> Error {
    Error::InvalidQuantization(format!("{}: {reason}", binding.source))
}

#[cfg(test)]
mod tests {
    use models::weights::{
        BindingTransform, ExpertProjectionRole, LayerTensorRole, LogicalTensorRole, TensorPacking,
    };

    use super::*;

    #[test]
    fn requires_typed_interleaved_expert_bank() -> Result<()> {
        let valid = binding(TensorPacking::InterleavedGateUp);
        let expected = BlockProjectionLayout::FusedGateUpBank { experts: 8, interleaved: true };
        assert_eq!(companions(&valid, expected)?, ("scales", "bias"));

        let invalid = binding(TensorPacking::Separate);
        assert!(companions(&invalid, expected).is_err());
        Ok(())
    }

    fn binding(packing: TensorPacking) -> TensorBinding {
        TensorBinding {
            role: LogicalTensorRole::Layer {
                index: 0,
                tensor: LayerTensorRole::ExpertProjection {
                    expert: None,
                    projection: ExpertProjectionRole::GateUp,
                },
            },
            source: "blocks".into(),
            shape: vec![8, 64, 1, 16],
            logical_shape: Some(vec![8, 64, 32]),
            transforms: vec![
                BindingTransform::StackedExperts { count: 8 },
                BindingTransform::FusedGateUp { interleaved: true },
            ],
            storage: TensorStorage::BlockQuantized {
                format: BlockQuantization::MXFP4,
                scales: "scales".into(),
                global_scale: None,
                input_scale: None,
                bias: Some("bias".into()),
                packing,
            },
        }
    }
}