decouple inference from streaming

set default interval to 3s

slightly increase repeat_penalty to reduce token repetition

bugfix on 'NoneType' object has no attribute 'caption'

update

add debug

update

initial commit