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

use mircuda::{DeviceBuffer, bf16};

use super::BatchedDecodeMoeBlockBf16;
use crate::{PagedDecodeBatch, Result, backend::block::graph::weights::CapturedBlockWeights};

/// One batched layer retaining immutable checkpoint weights.
pub struct BatchedDecodeMoeLayer {
    block: BatchedDecodeMoeBlockBf16,
    weights: CapturedBlockWeights,
}

impl BatchedDecodeMoeLayer {
    pub(in crate::backend::block) const fn new(
        block: BatchedDecodeMoeBlockBf16,
        weights: CapturedBlockWeights,
    ) -> Self {
        Self { block, weights }
    }

    pub fn execute(
        &mut self,
        input: &DeviceBuffer<bf16>,
        batch: &PagedDecodeBatch,
        output: &mut DeviceBuffer<bf16>,
    ) -> Result<()> {
        self.block.execute(input, self.weights.borrow(), batch, output)
    }
}