libmir-cuda 0.2.0

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
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144

use models::weights::{
    DecoderBoundaryBindings, RoutedDecoderLayerBindings, TensorBinding, TensorStorage,
};

use super::{
    ClampedRoutedConfig,
    layout::ClampedRoutedLayout,
    projection::{
        ClampedRoutedBoundaryProjection, ClampedRoutedLinearWeight, ClampedRoutedQkvWeight,
    },
    validation::validate_common,
};
use crate::{AffineQuantizedWeight, CudaBackend, CudaTensor, CudaTensorSet, Error, Result};

mod mlx;
mod native;

#[derive(Clone)]
pub(super) struct ClampedRoutedLayerWeights {
    pub input_norm: CudaTensor,
    pub qkv: ClampedRoutedQkvWeight,
    pub output: ClampedRoutedLinearWeight,
    pub output_bias: CudaTensor,
    pub sinks: CudaTensor,
    pub post_norm: CudaTensor,
    pub router: ClampedRoutedLinearWeight,
    pub router_bias: CudaTensor,
    pub experts: ClampedRoutedExpertWeights,
}

#[derive(Clone)]
pub(super) enum ClampedRoutedExpertWeights {
    Native(Box<NativeExpertWeights>),
    Mlx(Box<MlxExpertWeights>),
}

#[derive(Clone)]
pub(super) struct NativeExpertWeights {
    pub gate_up_blocks: CudaTensor,
    pub gate_up_scales: CudaTensor,
    pub gate_up_bias: CudaTensor,
    pub down_blocks: CudaTensor,
    pub down_scales: CudaTensor,
    pub down_bias: CudaTensor,
}

#[derive(Clone)]
pub(super) struct MlxExpertWeights {
    pub gate_blocks: CudaTensor,
    pub gate_scales: CudaTensor,
    pub gate_bias: CudaTensor,
    pub up_blocks: CudaTensor,
    pub up_scales: CudaTensor,
    pub up_bias: CudaTensor,
    pub down_blocks: CudaTensor,
    pub down_scales: CudaTensor,
    pub down_bias: CudaTensor,
}

pub(super) fn boundary(
    layout: ClampedRoutedLayout,
    tensors: &CudaTensorSet,
    bindings: DecoderBoundaryBindings<'_>,
    config: ClampedRoutedConfig,
) -> Result<(ClampedRoutedBoundaryProjection, CudaTensor, ClampedRoutedBoundaryProjection)> {
    let norm = tensor(tensors, &bindings.final_norm.source)?;
    match layout {
        ClampedRoutedLayout::Native => Ok((
            ClampedRoutedBoundaryProjection::Native(tensor(tensors, &bindings.embedding.source)?),
            norm,
            ClampedRoutedBoundaryProjection::Native(tensor(tensors, &bindings.output.source)?),
        )),
        ClampedRoutedLayout::Mlx => {
            let embedding = affine(tensors, bindings.embedding)?;
            let output = affine(tensors, bindings.output)?;
            embedding.infer_config(1, config.hidden, config.vocab)?;
            output.infer_config(1, config.hidden, config.vocab)?;
            Ok((
                ClampedRoutedBoundaryProjection::Mlx(embedding),
                norm,
                ClampedRoutedBoundaryProjection::Mlx(output),
            ))
        },
    }
}

fn affine(tensors: &CudaTensorSet, binding: &TensorBinding) -> Result<AffineQuantizedWeight> {
    AffineQuantizedWeight::load_binding(tensors, binding)
}

pub(super) fn layer(
    backend: &CudaBackend,
    layout: ClampedRoutedLayout,
    config: ClampedRoutedConfig,
    tensors: &CudaTensorSet,
    bindings: RoutedDecoderLayerBindings<'_>,
) -> Result<ClampedRoutedLayerWeights> {
    let input_norm = tensor(tensors, &bindings.input_norm.source)?;
    let q_bias = binding_bias(tensors, bindings.query)?;
    let k_bias = binding_bias(tensors, bindings.key)?;
    let v_bias = binding_bias(tensors, bindings.value)?;
    let output_bias = binding_bias(tensors, bindings.attention_output)?;
    let sinks = tensor(tensors, &bindings.attention_sinks.source)?;
    let post_norm = tensor(tensors, &bindings.post_attention_norm.source)?;
    let router_bias = binding_bias(tensors, bindings.router)?;
    validate_common(
        config,
        [&input_norm, &q_bias, &k_bias, &v_bias, &output_bias, &sinks, &post_norm, &router_bias],
    )?;
    let (qkv, output, router, experts) = match layout {
        ClampedRoutedLayout::Native => native::load(backend, config, tensors, bindings)?,
        ClampedRoutedLayout::Mlx => mlx::load(config, tensors, bindings)?,
    };
    Ok(ClampedRoutedLayerWeights {
        input_norm,
        qkv: ClampedRoutedQkvWeight {
            projections: qkv,
            biases: [q_bias, k_bias, v_bias],
        },
        output,
        output_bias,
        sinks,
        post_norm,
        router,
        router_bias,
        experts,
    })
}

fn binding_bias(tensors: &CudaTensorSet, binding: &TensorBinding) -> Result<CudaTensor> {
    let name = match &binding.storage {
        TensorStorage::Dense { bias, .. } | TensorStorage::BlockQuantized { bias, .. } => {
            bias.as_deref()
        },
        TensorStorage::AffineQuantized { output_bias, .. } => output_bias.as_deref(),
        TensorStorage::Auxiliary { .. } => None,
    }
    .ok_or_else(|| Error::MissingTensor(format!("bias for logical tensor {}", binding.source)))?;
    tensor(tensors, name)
}

pub(super) fn tensor(tensors: &CudaTensorSet, name: &str) -> Result<CudaTensor> {
    tensors.get(name).cloned().ok_or_else(|| Error::MissingTensor(name.into()))
}