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
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162

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

use super::{CudaBackend, NvFp4ExpertBank, grouped::GroupedNvFp4LinearBf16, view};
use crate::{
    Error, GatedActivation, Result,
    kernels::{
        ElementwiseBf16, NvFp4MicroBanks, NvFp4MicroGateLaunch, NvFp4MicroGateWorkspace,
        NvFp4MicroKernels, NvFp4MicroSpec,
    },
};

/// Direct gate/up followed by Tensor Core down projection and weighted
/// reduction.
#[derive(Debug)]
pub struct HybridNvFp4MoeBf16 {
    gate_kernels: NvFp4MicroKernels,
    gate: NvFp4ExpertBank,
    up: NvFp4ExpertBank,
    down_bank: NvFp4ExpertBank,
    down: GroupedNvFp4LinearBf16,
    reduce: ElementwiseBf16,
    gate_packed: DeviceBuffer<u8>,
    up_packed: DeviceBuffer<u8>,
    gate_scales: DeviceBuffer<u8>,
    up_scales: DeviceBuffer<u8>,
    down_output: DeviceBuffer<bf16>,
    stream: Stream,
    tokens: usize,
    selected: usize,
    hidden: usize,
}

impl CudaBackend {
    /// Prepares direct gate/up with a Tensor Core down tail for explicit gates.
    pub fn prepare_hybrid_gate_nvfp4_moe_bf16(
        &self,
        tokens: usize,
        selected: usize,
        activation: GatedActivation,
        gate: NvFp4ExpertBank,
        up: NvFp4ExpertBank,
        down: NvFp4ExpertBank,
    ) -> Result<HybridNvFp4MoeBf16> {
        HybridNvFp4MoeBf16::new(self, tokens, selected, activation, gate, up, down)
    }
}

impl HybridNvFp4MoeBf16 {
    #[allow(clippy::too_many_arguments)]
    fn new(
        backend: &CudaBackend,
        tokens: usize,
        selected: usize,
        activation: GatedActivation,
        gate: NvFp4ExpertBank,
        up: NvFp4ExpertBank,
        down_bank: NvFp4ExpertBank,
    ) -> Result<Self> {
        validate(&gate, &up, &down_bank)?;
        let experts = gate.config.experts;
        let hidden = gate.config.input_features;
        let intermediate = gate.config.output_features;
        let spec = NvFp4MicroSpec {
            experts,
            selected,
            hidden,
            intermediate,
            tokens,
            activation: activation.into(),
        };
        let groups = tokens
            .checked_mul(selected)
            .ok_or(Error::InvalidNvFp4("hybrid gate expert count overflow"))?;
        let down = GroupedNvFp4LinearBf16::new(backend, tokens, selected, down_bank.clone())?;
        let allocate_bytes =
            |count| backend.inner.pool.allocate::<u8>(&backend.inner.stream, count);
        Ok(Self {
            gate_kernels: NvFp4MicroKernels::compile(&backend.inner.compiler, spec)?,
            gate_packed: allocate_bytes(elements(groups, hidden / 2)?)?,
            up_packed: allocate_bytes(elements(groups, hidden / 2)?)?,
            gate_scales: allocate_bytes(elements(groups, hidden / 16)?)?,
            up_scales: allocate_bytes(elements(groups, hidden / 16)?)?,
            down_output: backend
                .inner
                .pool
                .allocate::<bf16>(&backend.inner.stream, elements(groups, hidden)?)?,
            reduce: ElementwiseBf16::compile(&backend.inner.compiler, hidden)?,
            stream: backend.inner.stream.clone(),
            gate,
            up,
            down_bank,
            down,
            tokens,
            selected,
            hidden,
        })
    }

    pub fn execute(
        &mut self,
        input: &DeviceBuffer<bf16>,
        selected: &DeviceBuffer<u32>,
        routing: &DeviceBuffer<bf16>,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        let scale_stride = self.down.scale_stride()?;
        {
            let (packed, scales) = self.down.quantized_workspace();
            self.gate_kernels.execute_gate(
                &self.stream,
                NvFp4MicroGateLaunch {
                    input,
                    selected,
                    banks: NvFp4MicroBanks {
                        gate: view(&self.gate),
                        up: view(&self.up),
                        down: view(&self.down_bank),
                    },
                    workspace: NvFp4MicroGateWorkspace {
                        gate_packed: &mut self.gate_packed,
                        up_packed: &mut self.up_packed,
                        gate_scales: &mut self.gate_scales,
                        up_scales: &mut self.up_scales,
                        output_packed: packed,
                        output_scales: scales,
                    },
                    output_scale_stride: scale_stride,
                },
            )?;
        }
        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 fn output_elements(&self) -> usize {
        self.tokens.saturating_mul(self.hidden)
    }
}

fn validate(gate: &NvFp4ExpertBank, up: &NvFp4ExpertBank, down: &NvFp4ExpertBank) -> Result<()> {
    let valid = gate.config == up.config
        && (gate.config.input_features, gate.config.output_features, gate.config.experts)
            == (down.config.output_features, down.config.input_features, down.config.experts);
    if valid {
        Ok(())
    } else {
        Err(Error::InvalidNvFp4("incompatible hybrid gate expert banks"))
    }
}

fn elements(rows: usize, columns: usize) -> Result<usize> {
    rows.checked_mul(columns)
        .ok_or(Error::InvalidNvFp4("hybrid gate workspace overflow"))
}