MoE-CAP
收藏资源简介:
MoE-CAP是一个专门用于评估稀疏混合专家系统(MoE)的基准数据集,由爱丁堡大学和微软研究院等机构创建。该数据集旨在评估MoE系统在成本、准确性和性能之间的权衡,包含新的稀疏感知性能指标和完整的部署成本模型。数据集的内容涵盖了多种MoE模型的性能评估,包括内存带宽利用率和模型FLOPS利用率等。创建过程涉及对现有MoE系统的广泛分析,并提出了新的评估方法。该数据集主要应用于大规模语言模型的优化和部署,旨在解决MoE系统在实际部署中的成本、准确性和性能之间的复杂权衡问题。
MoE-CAP is a benchmark dataset dedicated to evaluating sparse Mixture-of-Experts (MoE) systems, developed by institutions including the University of Edinburgh and Microsoft Research. This dataset aims to assess the trade-offs among cost, accuracy, and performance of MoE systems, and includes novel sparsity-aware performance metrics and a comprehensive deployment cost model. The dataset covers performance evaluations of various MoE models, such as memory bandwidth utilization and model FLOPS utilization. Its creation involves extensive analysis of existing MoE systems and the proposal of novel evaluation methodologies. This dataset is primarily applied to the optimization and deployment of large language models, and aims to address the complex trade-offs among cost, accuracy, and performance of MoE systems in real-world deployments.




