How to inference with buffer of frames?

Hi, I'm also working on mmaction recognition but I don't know how to do inference with a list of frames? Have you worked on this yet? If possible, can you give me some documentation so I can follow?