2000 character limit reached
Accelerating OpenPangu Inference on NPU via Speculative Decoding
Published 3 Mar 2026 in cs.DC | (2603.03383v1)
Abstract: To mitigate the Memory Wall bottleneck encountered by LLMs during inference on \textbf{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.
Paper Prompts
Sign up for free to create and run prompts on this paper using GPT-5.
Top Community Prompts
Collections
Sign up for free to add this paper to one or more collections.
