FP32 to INT8 Conversion | Dataford Interview Questions

Given an FP32 tensor and a floating-point scale factor, convert the input FP32 tensor to an INT8 tensor by dividing by the scale factor and casting to int8

Asked in the HM phone screen stage. Reported follow-up: how would asymmetric quantization change the conversion if a zero point were provided?

Contract

Implement convert_to_int8(tensor, scale). tensor is a non-empty, rectangular nested list of finite floating-point values. Divide each value by the positive scale, truncate toward zero, and saturate the result to the INT8 range [-128, 127]. Preserve the tensor's shape and return nested lists of Python integers.

Constraints

Tensor rank is at least 1.
All nested lists are rectangular.
The tensor contains at most 1000 floating-point elements.
Every tensor value is finite.
scale > 0.
Casting truncates toward zero.
Results are saturated to the inclusive INT8 range [-128, 127].

Function Signature

def convert_to_int8(tensor, scale):

Your question is FP32 to INT8 Conversion. Start with the requirements in the Question tab.

Run and submit as often as you like. When you're ready, talk me through your approach or go straight to the code.

Problem