SDSAT: Accelerating LLM Inference through Speculative Decoding with Semantic Adaptive Tokens | Synapse