像物理学家一样剪枝 LLM:把删 Block 变成 Ising 优化问题
Pruning LLMs Like a Physicist: Block Removal as an Ising Optimization Problem
Multiverse Computing 把“LLM 该删哪些层”写成 Ising 模型用物理方法解,比逐层打分剪枝的思路新鲜。
Multiverse Computing 的团队把 LLM 剪枝重写成一个 Ising 优化问题:Transformer 的每个 block 对应一个自旋,去留由系统能量最小化决定。
与按重要性分数逐层打分的常规剪枝不同,Ising 视角把“删哪些 block”当成一次全局组合优化来求解。文章给出了从构建 Ising 哈密顿量到实际删除 block 的完整推导,教程发布在 Hugging Face。
This report is categorized under AITOP's "Papers" desk, focusing on Europe's AI governance and technology ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include LLM, 模型剪枝, Multiverse Computing, Ising模型.
Primary source verification: published directly by Hugging Face: Blog with first-party authority.