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

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

use super::{CudaBackend, linear::Bf16Fp32Linear};
use crate::{
    CudaTensor, Error, Result,
    kernels::{RouterSpec, RouterTopK},
};

#[cfg(test)]
mod tests;

#[derive(Clone, Copy)]
pub struct RouterTensors<'a> {
    pub projection: &'a CudaTensor,
    pub norm_scale: &'a CudaTensor,
    pub expert_scale: &'a CudaTensor,
}

#[derive(Clone, Copy)]
pub struct RouterSelection<'a> {
    pub indices: &'a DeviceBuffer<u32>,
    pub weights: &'a DeviceBuffer<bf16>,
}

#[derive(Debug)]
pub struct RouterBf16 {
    operation: RouterTopK,
    projection: Bf16Fp32Linear,
    stream: Stream,
    normalized: DeviceBuffer<bf16>,
    scores: DeviceBuffer<f32>,
    selected: DeviceBuffer<u32>,
    weights: DeviceBuffer<bf16>,
    spec: RouterSpec,
    tokens: usize,
}

impl CudaBackend {
    pub fn prepare_router_bf16(&self, spec: RouterSpec) -> Result<RouterBf16> {
        RouterBf16::new(self, spec, 1)
    }

    pub(in crate::backend) fn prepare_router_batch_bf16(
        &self,
        spec: RouterSpec,
        tokens: usize,
    ) -> Result<RouterBf16> {
        RouterBf16::new(self, spec, tokens)
    }
}

impl RouterBf16 {
    fn new(backend: &CudaBackend, spec: RouterSpec, tokens: usize) -> Result<Self> {
        let selections = tokens
            .checked_mul(spec.top_k)
            .ok_or(Error::InvalidRouter("router batch size overflow"))?;
        let normalized_elements = tokens
            .checked_mul(spec.hidden)
            .ok_or(Error::InvalidRouter("router normalized size overflow"))?;
        let score_elements = tokens
            .checked_mul(spec.experts)
            .ok_or(Error::InvalidRouter("router score size overflow"))?;
        if tokens == 0 {
            return Err(Error::InvalidRouter("router batch is empty"));
        }
        Ok(Self {
            operation: RouterTopK::compile(&backend.inner.compiler, spec)?,
            projection: Bf16Fp32Linear::new(backend, tokens, spec.hidden, spec.experts)?,
            stream: backend.inner.stream.clone(),
            normalized: backend.inner.pool.allocate(&backend.inner.stream, normalized_elements)?,
            scores: backend.inner.pool.allocate(&backend.inner.stream, score_elements)?,
            selected: backend.inner.pool.allocate::<u32>(&backend.inner.stream, selections)?,
            weights: backend.inner.pool.allocate::<bf16>(&backend.inner.stream, selections)?,
            spec,
            tokens,
        })
    }

    pub fn execute(
        &mut self,
        input: &DeviceBuffer<bf16>,
        tensors: RouterTensors<'_>,
    ) -> Result<RouterSelection<'_>> {
        validate(tensors.projection, &[self.spec.experts, self.spec.hidden])?;
        validate(tensors.norm_scale, &[self.spec.hidden])?;
        validate(tensors.expert_scale, &[self.spec.experts])?;
        self.operation.normalize(
            &self.stream,
            input,
            bf16(tensors.norm_scale)?,
            &mut self.normalized,
            self.tokens,
        )?;
        self.projection
            .execute(&self.normalized, tensors.projection, &mut self.scores)?;
        self.operation.select(
            &self.stream,
            &self.scores,
            bf16(tensors.expert_scale)?,
            &mut self.selected,
            &mut self.weights,
            self.tokens,
        )?;
        Ok(RouterSelection {
            indices: &self.selected,
            weights: &self.weights,
        })
    }
}

fn validate(tensor: &CudaTensor, expected: &[usize]) -> Result<()> {
    if tensor.shape() == expected {
        Ok(())
    } else {
        Err(Error::InvalidQuantizedTensor {
            name: tensor.name().into(),
            expected: expected.to_vec(),
            actual: tensor.shape().to_vec(),
        })
    }
}

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