Efficient Knowledge Distillation for LLMs: Offline Top-K Logits and a Fused Chunked KL Loss ▲ 9 💬 2 2026-08-11 arXiv HF 原文 GitHub ★3 摘要 (无摘要)