📝 Speculate Deep and Accurate - Lossless and Training-Free Acceleration for Offloaded LLMs via Substitute Speculative Decoding