paperbot · PL 论文追踪

全部论文 待读 收藏

P2 Mugi: Value Level Parallelism For Efficient LLMs

Mugi：面向高效 LLM 的值级并行
ASPLOS Proceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 22026匹配度 1/10 · 论文聚焦 LLM 的硬件架构与值级并行、GEMM/量化优化，属于体系结构与高效推理方向；与用户关注的静态分析、形式化验证、浮点程序分析、贝叶斯程序分析以及 LLM/Agent 与 PL 的交叉均无直接关系。引用 1 inbox
Daniel Price, Prabhu Vellaisamy, John Paul Shen, Di Wu
语言与体系结构
crossref: 2026-09-12dblp: 2026-09-12arxiv: 2026-09-12

AI 速览

一句话：提出 Mugi 架构，将值级并行推广到 LLM 非线性与不对称小批量 GEMM，显著提升吞吐与能效。

已有值级并行（VLP）主要面向大批量、低精度的对称激活-权重 GEMM，难以覆盖 LLM 中更复杂的运算。本文探索 VLP 对 LLM 的收益：首先将 VLP 推广到非线性近似，采用以值为中心的策略，为重要数值分配更高精度，从而在端到端 LLM 精度、性能和效率上优于现有非线性近似；其次针对小批量、非对称输入的 GEMM 优化 VLP，并结合权重仅量化、KV cache 量化和分组查询注意力等 LLM 优化。在此基础上设计 Mugi 架构，支持完整 LLM 工作负载。实验表明，Mugi 对非线性 softmax 操作的吞吐和能效最高提升 45 倍与 668 倍，对 LLM 整体提升 2.07 倍与 3.11 倍，并将运营碳排降低 1.45 倍、隐含碳排降低 1.48 倍。

问题
现有值级并行仅针对大批量低精度对称激活-权重 GEMM，无法有效覆盖 LLM 中的非线性操作和小批量非对称 GEMM。
方法
将 VLP 推广到非线性近似，按数值重要性分配精度；针对小批量非对称 GEMM 优化 VLP，结合权重仅量化、KV cache 量化和分组查询注意力；设计 Mugi 架构以支持完整 LLM 工作负载。
评估
在 LLM 工作负载上测量吞吐、能效和碳排放，并与现有非线性近似及基线比较，报告了 softmax 非线性最高 45 倍/668 倍、整体 LLM 2.07 倍/3.11 倍的提升以及碳排降低。
原文摘要（Abstract）

Value level parallelism (VLP) has been proposed to improve the efficiency of large-batch, low-precision general matrix multiply (GEMM) between symmetric activations and weights. In transformer based large language models (LLMs), there exist more sophisticated operations beyond activation-weight GEMM. In this paper, we explore how VLP benefits LLMs. First, we generalize VLP for nonlinear approximations, outperforming existing nonlinear approximations in end-to-end LLM accuracy, performance, and efficiency. Our VLP approximation follows a value-centric approach, where important values are assigned with greater accuracy. Second, we optimize VLP for small-batch GEMMs with asymmetric inputs efficiently, which leverages timely LLM optimizations, including weight-only quantization, key-value (KV) cache quantization, and group query attention. Finally, we design a new VLP architecture, Mugi, to encapsulate the innovations above and support full LLM workloads, while providing better performance, efficiency and sustainability. Our experimental results show that Mugi can offer significant improvements on throughput and energy efficiency, up to 45× and 668× for nonlinear softmax operations, and 2.07× and 3.11× for LLMs, and also decrease operational carbon for LLM operation by 1.45× and embodied carbon by 1.48×.

链接与引用

DOI 原文 · arXiv · PDF（开放获取） · DBLP

BibTeX 
@inproceedings{PriceVS026,
  title = {Mugi: Value Level Parallelism For Efficient LLMs},
  author = {Daniel Price and Prabhu Vellaisamy and John Paul Shen and Di Wu},
  booktitle = {ASPLOS},
  year = {2026},
  doi = {10.1145/3779212.3790189}
}