GPTQ: Measuring Quantization Error
Intermediate
The problem, from first principles
Comparing quantized and original weights is the basic check of a quantization scheme. The mean absolute error summarizes how far the rounded weights are from the originals.
From theory to code
Implement quantization_error(x, bits), the mean absolute error of symmetric quantization.
Constraints
- Returns a float.
Hints
Hint 1
Quantize and dequantize as in the previous questions, then average the absolute differences.
Keywords
research-paperssystemsquantizationinference
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.