fn cuda_fwd()

in src/lib.rs [211:224]

• 14 lines of code
• 1 McCabe index (conditional complexity)

    fn cuda_fwd(
        &self,
        x: &candle::CudaStorage,
        x_l: &Layout,
    ) -> Result<(candle::CudaStorage, Shape)> {
        match x.dtype() {
            DType::F16 => self.fwd::<f16>(x, x_l, None, None),
            DType::BF16 => self.fwd::<bf16>(x, x_l, None, None),
            DType::F32 => self.fwd::<f32>(x, x_l, None, None),
            dt => {
                candle::bail!("fused-layer-norm is only supported for f32, f16 and bf16 ({dt:?})")
            }
        }
    }