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
142
143
144
145
146
147
148
149
150
151
152

use mircuda::{DeviceBuffer, Stream, bf16};

use super::{
    CudaBackend, NvFp4ExpertBank,
    grouped::{GroupedNvFp4LinearBf16, GroupedNvFp4PairBf16},
};
use crate::{Error, GatedActivation, MoeExecution, Result, kernels::ElementwiseBf16};

#[derive(Debug)]
enum ActivationStage {
    Split {
        operation: ElementwiseBf16,
        intermediate: DeviceBuffer<bf16>,
    },
    Fused,
}

/// Batched device-routed NVFP4 MLP using indexed grouped W4A4 CUTLASS.
#[derive(Debug)]
pub struct GroupedNvFp4MoeBf16 {
    gate_up: GroupedNvFp4PairBf16,
    down: GroupedNvFp4LinearBf16,
    activation_stage: ActivationStage,
    reduce: ElementwiseBf16,
    gate_output: DeviceBuffer<bf16>,
    up_output: DeviceBuffer<bf16>,
    down_output: DeviceBuffer<bf16>,
    activation: GatedActivation,
    stream: Stream,
    tokens: usize,
    selected: usize,
    output_elements: usize,
}

impl CudaBackend {
    #[allow(clippy::too_many_arguments)]
    pub(in crate::backend) fn prepare_grouped_nvfp4_moe_bf16(
        &self,
        tokens: usize,
        selected: usize,
        activation: GatedActivation,
        execution: MoeExecution,
        gate: NvFp4ExpertBank,
        up: NvFp4ExpertBank,
        down: NvFp4ExpertBank,
    ) -> Result<GroupedNvFp4MoeBf16> {
        GroupedNvFp4MoeBf16::new(self, tokens, selected, activation, execution, gate, up, down)
    }
}

impl GroupedNvFp4MoeBf16 {
    #[allow(clippy::too_many_arguments)]
    fn new(
        backend: &CudaBackend,
        tokens: usize,
        selected: usize,
        activation: GatedActivation,
        execution: MoeExecution,
        gate_bank: NvFp4ExpertBank,
        up_bank: NvFp4ExpertBank,
        down_bank: NvFp4ExpertBank,
    ) -> Result<Self> {
        let gate_up = GroupedNvFp4PairBf16::new(backend, tokens, selected, gate_bank, up_bank)?;
        let down = GroupedNvFp4LinearBf16::new(backend, tokens, selected, down_bank)?;
        let intermediate_elements = gate_up.output_elements()?;
        if down.output_features() == 0 || gate_up.output_features() != down.input_features() {
            return Err(Error::InvalidNvFp4("incompatible grouped expert banks"));
        }
        let down_elements = down.output_elements()?;
        let output_elements = tokens
            .checked_mul(down.output_features())
            .ok_or(Error::InvalidNvFp4("grouped MoE output overflow"))?;
        let allocate = |elements| backend.inner.pool.allocate(&backend.inner.stream, elements);
        let activation_stage = match execution {
            MoeExecution::IndexedGrouped => ActivationStage::Split {
                operation: ElementwiseBf16::compile(
                    &backend.inner.compiler,
                    intermediate_elements,
                )?,
                intermediate: allocate(intermediate_elements)?,
            },
            MoeExecution::FusedIndexedGrouped => ActivationStage::Fused,
            MoeExecution::DirectW4A4
            | MoeExecution::HybridW4A4
            | MoeExecution::SelectedWeightOnly
            | MoeExecution::Bucketed => {
                return Err(Error::InvalidExecutionPlan(
                    "decode MoE cannot use bucketed execution",
                ));
            },
        };
        Ok(Self {
            activation_stage,
            reduce: ElementwiseBf16::compile(&backend.inner.compiler, down.output_features())?,
            gate_output: allocate(intermediate_elements)?,
            up_output: allocate(intermediate_elements)?,
            down_output: allocate(down_elements)?,
            gate_up,
            down,
            activation,
            stream: backend.inner.stream.clone(),
            tokens,
            selected,
            output_elements,
        })
    }

    pub fn execute(
        &mut self,
        input: &DeviceBuffer<bf16>,
        selected: &DeviceBuffer<u32>,
        routing: &DeviceBuffer<bf16>,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        self.gate_up
            .execute(input, selected, &mut self.gate_output, &mut self.up_output)?;
        match &mut self.activation_stage {
            ActivationStage::Split { operation, intermediate } => {
                operation.gated(
                    &self.stream,
                    &self.gate_output,
                    &self.up_output,
                    intermediate,
                    self.activation.into(),
                )?;
                self.down.execute_ranked(intermediate, selected, &mut self.down_output)?;
            },
            ActivationStage::Fused => {
                self.down.quantize_gated(
                    &self.gate_output,
                    &self.up_output,
                    selected,
                    self.activation,
                )?;
                self.down.execute_prepared(selected, &mut self.down_output)?;
            },
        }
        self.reduce.weighted_reduce_batch(
            &self.stream,
            &self.down_output,
            routing,
            output,
            self.selected,
            self.tokens,
        )
    }

    #[must_use]
    pub const fn output_elements(&self) -> usize {
        self.output_elements
    }
}