libmir-cuda 0.3.0

CUDA 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 models::weights::{
    BlockProjectionLayout, BlockQuantization, RoutedDecoderLayerBindings, RoutedExpertBindings,
    TensorBinding, TensorStorage,
};

use super::{ClampedRoutedConfig, ClampedRoutedExpertWeights, NativeExpertWeights, tensor};
use crate::{
    CudaBackend, CudaTensorSet, Result,
    backend::clamped_routed::{
        projection::{ClampedRoutedLinearWeight, ClampedRoutedQkvProjections},
        validation::validate_native_experts,
    },
};

pub(super) fn load(
    backend: &CudaBackend,
    config: ClampedRoutedConfig,
    tensors: &CudaTensorSet,
    bindings: RoutedDecoderLayerBindings<'_>,
) -> Result<(
    ClampedRoutedQkvProjections,
    ClampedRoutedLinearWeight,
    ClampedRoutedLinearWeight,
    ClampedRoutedExpertWeights,
)> {
    let q = tensor(tensors, &bindings.query.source)?;
    let k = tensor(tensors, &bindings.key.source)?;
    let v = tensor(tensors, &bindings.value.source)?;
    let RoutedExpertBindings::InterleavedGateUp { gate_up, down } = bindings.experts else {
        return Err(crate::Error::InvalidDecoderKernel(
            "native clamped-routed requires interleaved gate/up expert bindings",
        ));
    };
    let (gate_up_scales, gate_up_bias) = block_companions(
        gate_up,
        BlockProjectionLayout::FusedGateUpBank {
            experts: config.experts,
            interleaved: true,
        },
    )?;
    let (down_scales, down_bias) =
        block_companions(down, BlockProjectionLayout::MatrixBank { matrices: config.experts })?;
    let experts = ClampedRoutedExpertWeights::Native(Box::new(NativeExpertWeights {
        gate_up_blocks: tensor(tensors, &gate_up.source)?,
        gate_up_scales: tensor(tensors, gate_up_scales)?,
        gate_up_bias: tensor(tensors, gate_up_bias)?,
        down_blocks: tensor(tensors, &down.source)?,
        down_scales: tensor(tensors, down_scales)?,
        down_bias: tensor(tensors, down_bias)?,
    }));
    validate_native_experts(config, &experts)?;
    Ok((
        ClampedRoutedQkvProjections::Native(backend.pack_bf16_linears([&q, &k, &v])?),
        ClampedRoutedLinearWeight::Native(tensor(tensors, &bindings.attention_output.source)?),
        ClampedRoutedLinearWeight::Native(tensor(tensors, &bindings.router.source)?),
        experts,
    ))
}

fn block_companions(
    binding: &TensorBinding,
    expected: BlockProjectionLayout,
) -> Result<(&str, &str)> {
    let TensorStorage::BlockQuantized {
        format: BlockQuantization::MXFP4,
        scales,
        bias: Some(bias),
        ..
    } = &binding.storage
    else {
        return Err(crate::Error::InvalidDecoderKernel(
            "native clamped-routed expert binding requires MXFP4 scales and bias",
        ));
    };
    if binding.block_projection_layout() != Some(expected) {
        return Err(crate::Error::InvalidDecoderKernel(
            "native clamped-routed expert binding has the wrong matrix-bank layout",
        ));
    }
    Ok((scales, bias))
}

#[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!(block_companions(&valid, expected)?, ("scales", "bias"));

        let invalid = binding(TensorPacking::Separate);
        assert!(block_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,
            },
        }
    }
}