Accelerating Production Llms With Combined Token/embedding Speculators

Wertheimer Davis, Rosenkranz Joshua, Parnell Thomas, Suneja Sahil, Ranganathan Pavithra, Ganti Raghu, Srivatsa Mudhakar. Arxiv 2024

[Paper]    
Training Techniques

This technical report describes the design and training of novel speculative decoding draft models, for accelerating the inference speeds of large language models in a production environment. By conditioning draft predictions on both context vectors and sampled tokens, we can train our speculators to efficiently predict high-quality n-grams, which the base model then accepts or rejects. This allows us to effectively predict multiple tokens per inference forward pass, accelerating wall-clock inference speeds of highly optimized base model implementations by a factor of 2-3x. We explore these initial results and describe next steps for further improvements.

Similar Work