This paper investigates the problem of question-answer generation for Vietnamese. Different from prior work that only focuses on question generation, we explore three tasks: question generation (QG), answer generation (AG), and question-answer generation (QAG) in a unified framework by using instruction tuning. To do that, we first define a set of instructions and then adapt this set to fine-tune QG, AG, and QAG models. Experimental results on five Vietnamese benchmark machine reading comprehension datasets show that instruction tuning is beneficial for task and domain adaptation of QG, QA, and QAG. We also introduce ViQAG, a system that offers the generation of the three tasks. Code and instructions are available here: https://github.com/Shaun-le/ViQAG .