libmir-cuda 0.1.0

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

use mircuda::{DeviceBuffer, bf16};
use models::{
    layout::{AttentionLayerType, DecoderConfig},
    weights::TensorCatalog,
};

use super::{NvFp4MoeLayerLoadConfig, model::payload_bytes, source::LayerSource};
use crate::{
    Bf16LinearPackWeights, Bf16LinearPairWeights, CudaBackend, CudaTensor, CudaTensorSet,
    DecodeAttentionOutputWeight, DecodeAttentionWeights, DecodeMoeBlockWeights,
    DecodeMoeLayerTemplate, DecodeQkvWeights, Error, NvFp4ExpertBankConfig, Result, RouterTensors,
};

impl CudaBackend {
    /// Loads one config-discovered NVFP4 routed-MoE layer into an immutable
    /// model-owned template.
    pub fn load_nvfp4_moe_layer_template(
        &self,
        decoder: &DecoderConfig,
        catalog: &TensorCatalog,
        layer: usize,
        load: NvFp4MoeLayerLoadConfig,
    ) -> Result<DecodeMoeLayerTemplate> {
        Ok(self.load_nvfp4_moe_layer_template_tracked(decoder, catalog, layer, load)?.0)
    }

    pub(super) fn load_nvfp4_moe_layer_template_tracked(
        &self,
        decoder: &DecoderConfig,
        catalog: &TensorCatalog,
        layer: usize,
        load: NvFp4MoeLayerLoadConfig,
    ) -> Result<(DecodeMoeLayerTemplate, u64)> {
        let block = load.block(decoder, layer)?;
        let key_is_value =
            decoder.attention_k_eq_v && decoder.layer_type(layer) == AttentionLayerType::Full;
        let source = LayerSource::discover(catalog, layer, block.experts, key_is_value)?;
        let tensors = upload(self, &source)?;
        let gate = self.prepare_nvfp4_expert_bank(
            bank(block.experts, block.attention.hidden_size, block.expert_intermediate),
            &source.gate,
        )?;
        let up = self.prepare_nvfp4_expert_bank(
            bank(block.experts, block.attention.hidden_size, block.expert_intermediate),
            &source.up,
        )?;
        let down = self.prepare_nvfp4_expert_bank(
            bank(block.experts, block.expert_intermediate, block.attention.hidden_size),
            &source.down,
        )?;
        let dense_gate_up = self.pack_bf16_linear_pair(
            tensor(&tensors, &source.names[9])?,
            tensor(&tensors, &source.names[10])?,
        )?;
        let qkv = self.pack_bf16_linears([
            tensor(&tensors, &source.names[1])?,
            tensor(&tensors, &source.names[2])?,
            tensor(&tensors, &source.names[3])?,
        ])?;
        let weights = weights(&tensors, &source.names, &qkv, &dense_gate_up)?;
        tracing::debug!(
            backend = "cuda",
            layer,
            prefix = source.prefix,
            tensors = source.tensors.len(),
            experts = block.experts,
            "loaded NVFP4 routed MoE layer template"
        );
        let bytes = source.payload_bytes()?;
        let template = self.prepare_decode_moe_layer_template(block, weights, gate, up, down)?;
        Ok((template, bytes))
    }
}

impl LayerSource<'_> {
    fn payload_bytes(&self) -> Result<u64> {
        let experts = self.gate.iter().chain(&self.up).chain(&self.down).flat_map(|source| {
            [source.weight, source.weight_scale, source.weight_scale_2, source.input_scale]
        });
        payload_bytes(self.tensors.iter().copied().chain(experts))
    }
}

fn upload(backend: &CudaBackend, source: &LayerSource<'_>) -> Result<CudaTensorSet> {
    let mut upload = backend.begin_tensor_upload();
    for tensor in &source.tensors {
        upload.enqueue(tensor)?;
    }
    upload.finish()
}

fn bank(experts: usize, input: usize, output: usize) -> NvFp4ExpertBankConfig {
    NvFp4ExpertBankConfig {
        experts,
        input_features: input,
        output_features: output,
    }
}

fn weights<'a>(
    tensors: &'a CudaTensorSet,
    names: &[String],
    qkv: &'a Bf16LinearPackWeights<3>,
    dense_gate_up: &'a Bf16LinearPairWeights,
) -> Result<DecodeMoeBlockWeights<'a>> {
    let get = |index: usize| tensor(tensors, &names[index]);
    Ok(DecodeMoeBlockWeights {
        attention: DecodeAttentionWeights {
            input_norm: get(0)?,
            qkv: DecodeQkvWeights::Bf16(qkv),
            query_norm: bf16_tensor(get(4)?)?,
            key_norm: bf16_tensor(get(5)?)?,
            output: DecodeAttentionOutputWeight::Bf16(get(6)?),
        },
        post_attention_norm: get(7)?,
        pre_dense_norm: get(8)?,
        dense_gate_up,
        dense_down: get(11)?,
        post_dense_norm: get(12)?,
        router: RouterTensors {
            projection: get(13)?,
            norm_scale: get(14)?,
            expert_scale: get(15)?,
        },
        pre_expert_norm: get(16)?,
        post_expert_norm: get(17)?,
        post_feed_forward_norm: get(18)?,
        layer_scalar: get(19)?,
    })
}

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

fn bf16_tensor(tensor: &CudaTensor) -> Result<&DeviceBuffer<bf16>> {
    tensor.as_bf16().ok_or_else(|| Error::DTypeMismatch {
        name: tensor.name().into(),
        expected: "BF16",
    })
}