Struct QuantizePerTokenBackwardArgs 

pub struct QuantizePerTokenBackwardArgs<'a, TIn: Element> {
    pub d_output: TensorRef<'a, TIn, 2>,
    pub input: TensorRef<'a, TIn, 2>,
    pub scale: TensorRef<'a, TIn, 1>,
    pub zero_point: TensorRef<'a, i32, 1>,
    pub d_input: TensorMut<'a, TIn, 2>,
}

Args bundle for the per-token BW launch.

Fields

d_output: TensorRef<'a, TIn, 2>

Upstream gradient [N, D].

input: TensorRef<'a, TIn, 2>

Saved input from FW (needed for the in-range mask) [N, D].

scale: TensorRef<'a, TIn, 1>

Saved scale [N].

zero_point: TensorRef<'a, i32, 1>

Saved zero-point [N].

d_input: TensorMut<'a, TIn, 2>

Output dx [N, D].