Fast Chain-of-Thought: A Glance of Future from Parallel Decoding Leads to Answers Faster

In this work, we propose FastCoT, a model-agnostic framework based onparallel decoding without any further training of an auxiliary model ormodification to the LLM itself. FastCoT uses a size-varying context windowwhose size changes with position to conduct parallel decoding andauto-regressive decoding simultaneously, thus fully utilizing GPU computationresources. In FastCoT, the parallel decoding part provides the LLM with a quickglance of the future composed of approximate tokens, which could lead to fasteranswers compared to regular autoregressive decoding used by causaltransformers. We also provide an implementation of parallel decoding withinLLM, which supports KV-cache generation and batch processing. Through extensiveexperiments, we demonstrate that FastCoT saves inference time by nearly 20with only a negligible performance drop compared to the regular approach.Additionally, we show that the context window size exhibits considerablerobustness for different tasks.