run GPT-OSS 20B and Qwen3 30B A3B with intel arrow lake igpu on vulkan and sycl backend ,the best decode speed can be 20 token per second, but the prefill also only 20+ token/s, I have no idea about it