Break the Sequential Dependency of LLM Inference Using Lookahead Decoding

Lookahead decoding: exact, parallel decoding algorithm requiring no auxiliary models or data stores. Trades per-step FLOPs to reduce total decoding steps. Up to 1.8x speedup on MT-bench, 4x with strong scaling on multiple GPUs in code completion. Compatible with FlashAttention. Uses Jacobi iteration to look ahead and decode multiple tokens in parallel.