sampling 1 Speculative Decoding - Making Language Models Generate Faster Without Losing Their Minds Apr 21, 2025