Accelerating OpenPangu Inference on NPU via Speculative Decoding

To mitigate the Memory Wall bottleneck encountered by Large Language Models (LLMs) during inference on $\mathbf{\text{NPU}}$ hardware, and addressing the scarcity of native support for mainstream speculative decoding algorithms on domestic infrastructure, this study presents an end-to-end speculative inference acceleration scheme for OpenPangu-7B.