Your LLM Is Guessing Ahead. Then It Checks Itself aka Speculative Decoding

Every token your LLM generates costs one full forward pass. One pass, one token. No shortcuts.