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

use mircuda::{DeviceBuffer, Stream, VariableGroupedFp4Spec, bf16};

use super::{CudaBackend, NvFp4ExpertBank, moe::ExpertBuckets};
use crate::{
    Error, Result,
    kernels::{BucketQuantize, NvFp4BucketPreparation, NvFp4Spec, scale_elements},
};

#[derive(Debug)]
pub(super) struct BucketedNvFp4Projection {
    pub(super) bank: NvFp4ExpertBank,
    pub(super) packed: DeviceBuffer<u8>,
    pub(super) scales: DeviceBuffer<u8>,
    pub(super) stream: Stream,
    spec: NvFp4Spec,
    assignments: usize,
    experts: usize,
    tokens: usize,
    selected: usize,
}

impl BucketedNvFp4Projection {
    pub(super) fn new(
        backend: &CudaBackend,
        tokens: usize,
        selected: usize,
        bank: NvFp4ExpertBank,
    ) -> Result<Self> {
        let assignments = tokens
            .checked_mul(selected)
            .ok_or(Error::InvalidNvFp4("bucketed assignment count overflow"))?;
        let experts = bank.config.experts;
        if tokens == 0 || selected == 0 || selected > experts {
            return Err(Error::InvalidNvFp4("invalid bucketed expert geometry"));
        }
        let spec = NvFp4Spec::new(bank.config.input_features, bank.config.output_features)?;
        let packed_elements = assignments
            .checked_mul(spec.input_features / 2)
            .ok_or(Error::InvalidNvFp4("bucketed packed input overflow"))?;
        let scale_count = experts
            .checked_mul(scale_elements(assignments, spec.input_features)?)
            .ok_or(Error::InvalidNvFp4("bucketed input scale overflow"))?;
        Ok(Self {
            bank,
            packed: backend.inner.pool.allocate(&backend.inner.stream, packed_elements)?,
            scales: backend.inner.pool.allocate_zeroed(&backend.inner.stream, scale_count)?,
            stream: backend.inner.stream.clone(),
            spec,
            assignments,
            experts,
            tokens,
            selected,
        })
    }

    pub(super) fn quantize_pair(
        preparation: &NvFp4BucketPreparation,
        buckets: &ExpertBuckets,
        input: &DeviceBuffer<bf16>,
        selected: &DeviceBuffer<u32>,
        left: &mut Self,
        right: &mut Self,
    ) -> Result<()> {
        left.require_compatible(right)?;
        let geometry = left.quantize_geometry(false);
        preparation.quantize_pair(
            &left.stream,
            input,
            selected,
            &buckets.order,
            &buckets.offsets,
            &left.bank.input_scales,
            &right.bank.input_scales,
            &mut left.packed,
            &mut right.packed,
            &mut left.scales,
            &mut right.scales,
            geometry,
        )
    }

    pub(super) fn quantize_ranked(
        &mut self,
        preparation: &NvFp4BucketPreparation,
        buckets: &ExpertBuckets,
        input: &DeviceBuffer<bf16>,
        selected: &DeviceBuffer<u32>,
    ) -> Result<()> {
        let geometry = self.quantize_geometry(true);
        preparation.quantize(
            &self.stream,
            input,
            selected,
            &buckets.order,
            &buckets.offsets,
            &self.bank.input_scales,
            &mut self.packed,
            &mut self.scales,
            geometry,
        )
    }

    pub(super) fn plan_spec(&self) -> Result<VariableGroupedFp4Spec> {
        Ok(VariableGroupedFp4Spec::new(
            self.experts,
            self.experts,
            self.assignments,
            self.spec.output_features,
            self.spec.input_features,
            self.assignments,
        )?)
    }

    pub(super) fn output_elements(&self) -> Result<usize> {
        self.assignments
            .checked_mul(self.spec.output_features)
            .ok_or(Error::InvalidNvFp4("bucketed expert output overflow"))
    }

    pub(super) const fn output_features(&self) -> usize {
        self.spec.output_features
    }

    pub(super) const fn input_features(&self) -> usize {
        self.spec.input_features
    }

    fn require_compatible(&self, other: &Self) -> Result<()> {
        if self.spec.input_features == other.spec.input_features
            && self.spec.output_features == other.spec.output_features
            && self.assignments == other.assignments
            && self.experts == other.experts
            && self.tokens == other.tokens
            && self.selected == other.selected
        {
            Ok(())
        } else {
            Err(Error::InvalidNvFp4("incompatible paired bucket projections"))
        }
    }

    const fn quantize_geometry(&self, ranked: bool) -> BucketQuantize {
        BucketQuantize {
            assignments: self.assignments,
            experts: self.experts,
            selected: self.selected,
            input_rows: if ranked {
                self.assignments
            } else {
                self.tokens
            },
            columns: self.spec.input_features,
            ranked,
        }
    }
}