Faster without a draft model? Measuring vLLM's Speculative Decoding (n-gram method)

Speculative Decoding is a technique that speeds up generation by having a lightweight "draft model" predict several tokens ...