Efficient LLM inference solution on Intel GPU