What are the trade-offs of using post-training quantization such as INT8 or FP4 on model accuracy for code generation tasks?
