lucidrains/speculative-decoding ? reverse-engineered prompt

Build me a Python project that explores faster transformer decoding with speculative decoding. I want it to include a basic speculative decoding setup, plus the extra ideas mentioned in the README like early exit and a ProphetNet style future prediction approach, so I can compare which one works best.

It should be easy to train and test these variants on small language modeling style data, and have scripts that let me run experiments, measure speed, and check that the outputs still look correct. If you can, make it practical for batched generation too, since different sequences may finish at different times. I’d also like the code to be organized so it’s easy to plug into a larger text generation system later, especially something like a text to semantic decoder. If you need to look up current docs or best practices for any of the model tricks, feel free to do that.