speculative sampling

**Speculative Sampling** is **the sampling technique that generates high-quality samples from language models faster by using approximate sampling methods with verification** — achieving 1.5-3× speedup for sampling-based generation while maintaining exact output distribution, enabling faster creative text generation, diverse outputs, and efficient exploration of model capabilities. **Sampling in Language Models:** - **Autoregressive Sampling**: at each step, sample token from P(x_t | x_{