Worked on the pytorch/ao repository to stabilize the int8 dynamic activation quantization path following the v2 tensor migration. Addressed an accuracy regression by introducing default handling for scale_dtype, preserving precision with float32 and adding configurability through QuantizeTensorToInt8Kwargs. The solution involved threading the scale_dtype parameter through Int8Tensor.from_hp() and simplifying the quantization logic by hardcoding scale_dtype to float32 in the choose_qparams_affine path, thereby maintaining previous behavior and preventing unnecessary API expansion. Utilized Python and machine learning techniques with a focus on quantization, ensuring cross-platform reliability and precision for quantized inference workflows in the codebase.