use super::RudaTensor;
use crate::dsl::Runtime;
use super::allocation::empty_device_dtype;
use ruda_core::tensor::{DType, TensorMetadata};
pub fn dequantize<R>(tensor: RudaTensor<R>, dtype: DType) -> RudaTensor<R>
where
R: Runtime,
{
let scheme = match tensor.dtype {
DType::QFloat(scheme) => scheme,
_ => return tensor,
};
let output = empty_device_dtype(
tensor.client.clone(),
tensor.device.clone(),
tensor.shape(),
dtype,
);
let (values, params) = tensor.quantized_handles().unwrap();
crate::quantization::dequantize::launch_ref(
&output.client,
values.binding(),
output.clone().binding(),
params.binding(),
&scheme,
dtype.into(),
)
.expect("Kernel to never fail");
output
}