There is nothing conflicting between the two, so you can call caffe routines from within Spark, while keeping in mind all constraints the latter puts on you. How would you train the model? I guess in your question you are asking if there are specific bindings that allow that. Not to my knowledge. If you are bent on Spark, try with
H2O library - it's nicely integrated with Spark. With caffe I would stick to GPUs.
Naturally I can be completely and utterly wrong, it's my first day with caffe and I am still trying to make heads or tails with it...
Cheers,
Lucas