The open-source inferencing software increases throughput and reduces the cost of LLM token generation, the chipmaker said.