/ The Arena

Problem Overview

Speculative Decoding achieves 2x-3x speedups by using a fast draft model to propose $K$ tokens, verified in parallel by the target model via Modified Rejection Sampling.

In this capstone, you will implement the deterministic Speculative Verification Engine that:

1. Validates candidate draft tokens against target distribution ratios $P(x)/Q(x)$.
2. Accepts valid tokens and resamples rejected positions from normalized residual distributions $(P-Q{)}^{+}$.
3. Emits bonus tokens when all $K$ candidates are accepted.