Large language models spend most decoding time waiting to read weights from memory. One token at a time keeps the process memory-bound on GPUs. Speculative decoding lets a small drafter model propose several tokens at once. The target model verifies them in a single pass and accepts the matching prefix. Greedy decoding uses exact matching of the most likely token. Speculative sampling extends this to temperature-based sampling. Gumbel coupling shares noise between drafter and target for reproducible results. Acceptance rate measures how often drafter guesses are kept. The method speeds up decoding without changing the output distribution. Industry-standard drafters include EAGLE-3, multi-token predictors, and block-diffusion models like DFlash and DSpark. Training a drafter requires minimizing distribution mismatch, not just KL divergence.

• ≤208 tokens/s for 8B model at batch size 1
• 295 tokens to break memory-bound limit
• 1 - TV(p,q) acceptance rate bound