Open tmostak opened 1 month ago
Implementing multi-token prediction in NanoGPT could be very amazing it would increase the sample efficiency so we can get more accurate results as well as models can perform speculative decoding, where they generate multiple possible sequences of tokens and then we can choose the most likely one
Was anybody able to do this? I think we can add this functionality.
Per the recent paper from Meta, it appears that models that predict multiple future tokens can exhibit significantly greater sample efficiency than models trained only on next-token prediction, plus the extra token heads can be used to implement speculative decoding to speed up inference (up to 3X in their experiments), without the need for a draft model.
It would be amazing to see multi-token prediction implemented in nanoGPT, as it would allow the community to easily experiment with this promising technique.