Faster LLM decoding for free: how EAGLE-3 speeds up generation without changing a single output

Large language models write one token at a time. To produce a word, the model reads everything so far and runs a full forward pass to pick the next token, then does it again, and again. That sounds li