Accelerating Production LLMs with Combined Token/Embedding Speculators