A dynamic weight precision allocation strategy is proposed, utilizing low-bit quantization and uncertainty-based gating to selectively upgrade weig...
论文概览
- arXiv ID:
2611.14418 - 发布机构 / 来源:Hugging Face Daily Papers
- 作者:Mohammed Nayeem Khan, Logesh L, Aliyas Ahmad Misri
- 社区热度:4 Upvotes
核心摘要(Abstract / AI Summary)
A dynamic weight precision allocation strategy is proposed, utilizing low-bit quantization and uncertainty-based gating to selectively upgrade weight precision during LLM generation, thereby optimizing computational cost.
