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

mod pair;
mod projection;

use mircuda::{DeviceBuffer, VariableGroupedFp4Plan, bf16};
pub(super) use pair::BucketedNvFp4PairBf16;
use projection::BucketedNvFp4Projection;

use super::{CudaBackend, NvFp4ExpertBank, bucketed_moe::ExpertBuckets};
use crate::{Error, Result, kernels::NvFp4BucketPreparation};

/// Expert-bucketed W4A4 projection with variable rows per CUTLASS group.
#[derive(Debug)]
pub(super) struct BucketedNvFp4LinearBf16 {
    plan: VariableGroupedFp4Plan,
    projection: BucketedNvFp4Projection,
}

impl BucketedNvFp4LinearBf16 {
    pub(super) fn new(
        backend: &CudaBackend,
        tokens: usize,
        selected: usize,
        bank: NvFp4ExpertBank,
    ) -> Result<Self> {
        let projection = BucketedNvFp4Projection::new(backend, tokens, selected, bank)?;
        let plan = VariableGroupedFp4Plan::new(
            &backend.inner.context,
            &backend.inner.stream,
            projection.plan_spec()?,
        )?;
        Ok(Self { plan, projection })
    }

    pub(super) fn execute_ranked(
        &mut self,
        preparation: &NvFp4BucketPreparation,
        buckets: &ExpertBuckets,
        input: &DeviceBuffer<bf16>,
        selected: &DeviceBuffer<u32>,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        self.projection.quantize_ranked(preparation, buckets, input, selected)?;
        self.execute_prepared(buckets, output)
    }

    fn execute_prepared(
        &mut self,
        buckets: &ExpertBuckets,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        validate_output(&self.projection, output)?;
        let projection = &self.projection;
        Ok(self.plan.execute(
            &projection.stream,
            &projection.packed,
            &projection.scales,
            &projection.bank.weight,
            &projection.bank.cutlass_scales,
            &projection.bank.combined_scales,
            &buckets.indices,
            &buckets.counts,
            &buckets.offsets,
            output,
        )?)
    }

    pub(super) fn output_elements(&self) -> Result<usize> {
        self.projection.output_elements()
    }

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

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

fn validate_output(
    projection: &BucketedNvFp4Projection,
    output: &DeviceBuffer<bf16>,
) -> Result<()> {
    let expected = projection.output_elements()?;
    if output.len() == expected {
        Ok(())
    } else {
        Err(Error::QuantizedGemvLengthMismatch {
            operand: "bucketed NVFP4 output",
            expected,
            actual: output.len(),
        })
    }
}