With spot instances you just have to be prepared for them to get terminated at random. If you're processing messages off a queue in most cases you get this for free - the ack deadline will expire and the message will be redelivered to another subscriber.

Docker container might be a good shout, though I've not done anything using a GPU from docker before