TR-PTQ: High-Accuracy Integer-Only Transformer Post Training Quantization via Taylor Region Reformulation
The research addresses accuracy loss during post-training quantization of transformer nonlinearities by identifying structural error sources, not just numerical precision issues. It proposes TR-PTQ, a unified integer-only framework using Taylor Region approximations for exponential and logarithm functions. This method allows complex operations to run on integer hardware while achieving high accuracy through targeted parameter optimization.