[ICLR 2025] SWIFT: On-the-Fly Self-Speculative Decoding for LLM Inference Acceleration - View it on GitHub