Struct DynamicRangeQuantizeArgs 

pub struct DynamicRangeQuantizeArgs<'a, TIn: Element, TOut: IntElement> {
    pub input: TensorRef<'a, TIn, 2>,
    pub scale_out: TensorMut<'a, TIn, 1>,
    pub output: TensorMut<'a, TOut, 2>,
}

Args bundle for a dynamic_range_quantize launch.

Compared with super::QuantizePerTokenArgs, the caller does NOT supply scale / zero_point — those are computed by the kernel from the runtime dynamic range. The plan writes scale[N] into the caller-supplied scale_out buffer so a downstream dequantize step has access to the same scale.

zero_point is implicit (= 0 for symmetric) and is not materialized.

Fields

input: TensorRef<'a, TIn, 2>

Input [N, D] in FP.

scale_out: TensorMut<'a, TIn, 1>

Per-row scale [N] in FP — written by the kernel.

output: TensorMut<'a, TOut, 2>

Output [N, D] in int.