causal-conditioning

JetSpec is a speculative decoding framework that combines efficient forward drafting with causal conditioning to improve LLM inference speed and acceptance rates, achieving up to 9.64x speedup on MATH-500 and 4.58x on conversational workloads.