Struct QuantizePerTensorBackwardArgs 

pub struct QuantizePerTensorBackwardArgs<'a, TIn: Element, TOut: IntElement> {
    pub input: TensorRef<'a, TIn, 1>,
    pub scale: <TIn as Element>::Scalar,
    pub zero_point: i32,
    pub d_output: TensorRef<'a, TIn, 1>,
    pub d_input: TensorMut<'a, TIn, 1>,
    pub _phantom: PhantomData<TOut>,
}

Args bundle for a quantize_per_tensor backward launch.

Fields

input: TensorRef<'a, TIn, 1>

Saved FW input [numel] in FP — required for mask recomputation.

scale: <TIn as Element>::Scalar

Scalar scale (FP, same value used in FW).

zero_point: i32

Scalar zero point (same value used in FW).

d_output: TensorRef<'a, TIn, 1>

Upstream gradient [numel] in FP.

d_input: TensorMut<'a, TIn, 1>

Output [numel] in FP — same dtype as d_output.

_phantom: PhantomData<TOut>

Phantom for the int-output dtype carried by the plan type parameter (so the BW plan has the same <TIn, TOut> shape as the FW plan, even though the BW kernel doesn’t consume an int operand).