vllm

507df79a - [Hybrid] Simplify accepted token counting in spec decode for hybrid models (#38372)