AI Research & Engineering: RecSys, Search, NLP, Generative AI and Beyond

Tag DeepSeekMoE

Auxiliary-Loss-Free Load Balancing 详解:用 bias 替代 balance loss,消除 MoE 训练的隐性梯度污染(DeepSeek 系列第 9 篇)

Auxiliary-Loss-Free (arXiv:2408.15664) 详解:用 expert-wise bias + 规则式更新替代传统 auxiliary balance loss,消除干扰梯度对主任务训练的污染。balance 与 specialization 通过 bias 与 affinity 解耦。V3 训练全面采纳。

Loading

ESFT 详解:只更新任务相关 expert,让 MoE 模型的 fine-tuning 成本降低 90%(DeepSeek 系列第 8 篇)

ESFT (Expert-Specialized Fine-Tuning, arXiv:2407.01906) 详解:基于 MoE 模型在下游任务上 expert 激活的天然稀疏性,只 fine-tune 任务相关的少数 expert,5-25% 可训练参数即可匹配 Full FT 性能,明显优于 LoRA。

Loading

DeepSeekMoE 详解:Fine-grained Expert 与 Shared Expert 双柱设计的奠基之作(DeepSeek 系列第 2 篇)

转载本文请注明出处:https://yudonglee.me/deepseekm… Continue Reading →

Loading

© 2026 Yudong‘s Blog — Powered by WordPress

Theme by Anders NorenUp ↑