Int8 to Float32 Reconstruction (Dequantize)

The problem, from first principles

Storing weights as int8 is only useful if they can be turned back into something a floating-point computation can actually use. 01-quantize-float32-to-int8's scale/zero_point are exactly the two numbers needed to invert that mapping — but the inversion is necessarily lossy: rounding to the nearest integer during quantization permanently discards whatever fraction of a "step" the true value didn't land exactly on.

From theory to code

Implement dequantize(q, scale, zero_point), which inverts quantize's forward mapping, and quantization_error(x, q, scale, zero_point), which measures how much information that round trip actually lost.

Constraints

• dequantize(q, scale, zero_point) returns a float array the same shape as q.
• quantization_error returns a single float: the maximum absolute difference between the original values and their reconstruction.
• Casting q to a floating-point dtype must happen before subtracting zero_point — subtracting first, while still an int8 array, can silently wrap around on overflow.

Hints

Open one at a time. Each gives away a little more than the last.