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
129
130
131
132
133
134
135
136
137

use std::collections::HashSet;

use models::{
    layout::DecoderConfig,
    semantic::SemanticModelSpec,
    weights::{
        LayerTensorRole, LogicalTensorRole, TensorCatalog, TensorInfo, TensorStorage,
        WeightBindingPlan,
    },
};

use super::{SharedRoutedModelLoadConfig, model::payload_bytes};
use crate::{CudaBackend, CudaSharedRoutedModelTemplate, Error, Result};

impl CudaBackend {
    pub fn load_shared_routed_model_template(
        &self,
        decoder: &DecoderConfig,
        catalog: &TensorCatalog,
        load: SharedRoutedModelLoadConfig,
    ) -> Result<CudaSharedRoutedModelTemplate> {
        let mut ignored = |_completed, _detail| {};
        let spec = SemanticModelSpec::discover(decoder, catalog)?;
        let bindings = WeightBindingPlan::discover(&spec, catalog)?;
        self.load_shared_routed_model_template_with_progress(
            decoder, &spec, catalog, &bindings, load, &mut ignored,
        )
    }

    pub(crate) fn load_shared_routed_model_template_with_progress(
        &self,
        decoder: &DecoderConfig,
        semantic: &SemanticModelSpec,
        catalog: &TensorCatalog,
        bindings: &WeightBindingPlan,
        load: SharedRoutedModelLoadConfig,
        progress: &mut dyn FnMut(u64, String),
    ) -> Result<CudaSharedRoutedModelTemplate> {
        if load.max_sequence_blocks == 0 {
            return Err(Error::UnsupportedDecoderLayer(
                "shared-routed model sequence block capacity is empty".into(),
            ));
        }
        let source = shared_routed_source(bindings, decoder.num_hidden_layers, catalog)?;
        let raw = raw_sources(bindings);
        let cast = self.prepare_dense_cast()?;
        let mut upload = self.begin_tensor_upload();
        for tensor in &source {
            if raw.contains(tensor.name.as_str()) {
                upload.enqueue(tensor)?;
            } else {
                upload.enqueue_float_as_bf16(tensor, &cast)?;
            }
        }
        let tensors = upload.finish()?;
        let bytes = payload_bytes(source.iter().copied())?;
        progress(bytes, format!("uploaded {} shared-routed checkpoint tensors", source.len()));
        tracing::debug!(
            backend = "cuda",
            layers = decoder.num_hidden_layers,
            tensors = source.len(),
            bytes,
            "loaded shared-routed mixed-mixer model template"
        );
        CudaSharedRoutedModelTemplate::from_tensors(
            self,
            decoder,
            semantic,
            &tensors,
            catalog,
            bindings,
            load.cache,
            load.max_sequence_blocks,
        )
    }
}

fn shared_routed_source<'a>(
    bindings: &WeightBindingPlan,
    layers: usize,
    catalog: &'a TensorCatalog,
) -> Result<Vec<&'a TensorInfo>> {
    let individual = bindings
        .tensors
        .iter()
        .filter(|binding| {
            matches!(
                binding.role,
                LogicalTensorRole::Layer {
                    tensor: LayerTensorRole::ExpertProjection { expert: Some(_), .. },
                    ..
                }
            )
        })
        .flat_map(|binding| binding.physical_sources())
        .collect::<HashSet<_>>();
    let mut names = bindings.decoder_boundary()?.physical_sources();
    for layer in 0..layers {
        names.extend(bindings.hybrid_decoder_layer(layer)?.physical_sources());
    }
    let mut seen = HashSet::new();
    names
        .into_iter()
        .filter(|name| !individual.contains(name))
        .filter(|name| seen.insert(*name))
        .map(|name| required(catalog, name))
        .collect()
}

fn raw_sources(bindings: &WeightBindingPlan) -> HashSet<&str> {
    bindings
        .tensors
        .iter()
        .filter(|binding| {
            matches!(
                binding.storage,
                TensorStorage::AffineQuantized { .. }
                    | TensorStorage::BlockQuantized { .. }
                    | TensorStorage::Float8 { .. }
                    | TensorStorage::PackedInt8 { .. }
                    | TensorStorage::PackedInt4 { .. }
                    | TensorStorage::Awq { .. }
                    | TensorStorage::Gptq { .. }
                    | TensorStorage::BitsAndBytes4Bit { .. }
            )
        })
        .flat_map(|binding| binding.physical_sources())
        .collect()
}

fn required<'a>(catalog: &'a TensorCatalog, name: &str) -> Result<&'a TensorInfo> {
    catalog
        .tensors
        .iter()
        .find(|tensor| tensor.name == name)
        .ok_or_else(|| Error::MissingTensor(name.into()))
}