4 ms·Cassandra: Enabling Reasoning LLMs at Edge via Self-Speculative Decoding4 points by chrsw 5mo ago