Hugging Face Trending Papers

DraftExpert: Expansion-Aware Self-Speculative Decoding for End-Device MoE Inference

Read the original on Hugging Face Trending Papers →

Large Mixture-of-Experts (MoE) language models are attractive for end-device deployment because only a small subset of experts is active per token, but their routed expert weights often exceed accelerator memory. We target latency-critical single-user settings where routed experts are staged on demand from CPU memory to a GPU or from Flash to a mobile NPU.

Summary generated by The Flow from the publisher's feed. The full article lives at Hugging Face Trending Papers.