新开一个系列连载,小周带你读论文,会不定期的更新各种新的,甚至老的有价值的论文,当然您有时间自己读最好了,如果自己读嫌麻烦,可以来看我这个的总结
老规矩,1,2,3 上链接...
IEIT-Yuan/Yuan-2.0: Yuan 2.0 Large Language Model (github.com)
Yuan2是浪潮的刚发布的LLM是基于Yuan1改的(这里吐槽一下浪潮,Yuan1的pretrain数据原来是公开下载的有1T多的语料很大一部分中文比例,现在给关闭了

)
Yuan2这论文写的还是有点意思的,受限于算力要求,很多事实性的实验我没法做证明或者证伪,那就先看看文中的一些理论创新
1- 魔改Transformer(LFA):

为了好理解我沾个Llama2的结构作为对比

几乎一眼就可以看出来变化,他把multiheader attention层给改了(其实要严格一点说也不算全改,只是前面加东西了)!Transformer玩的啥呢,其实就是玩attetion这层呢,他为什么要把核心内容给改了呢?
下面是论文里给的说法:
Attention, as a basic building block in LLMs, has showed great success across NLP tasks [9,10]. When a sequence is fed in to a language model, attention mechanism learns the weights of each pair of tokens to build the dependencies across the entire input sequence. The mechanism equally treats a token in neighbourhood and that in a distance. However, in natural language, the dependencies of words in neighbourhood are often stronger than the words faraway. The interconnection learned by Attentio

本文是小周带你读论文系列的第一篇,主要探讨了浪潮Yuan2在Transformer结构的改进——局部过滤注意力(LFA)和分布式训练时间预测法。LFA通过预处理QKV输入,引入局部依赖。此外,文章还介绍了非均匀PP(Non-uniform PP)解决显存分配不均的问题。尽管存在争议,Yuan2的创新对于理解LLM优化具有参考价值。
最低0.47元/天 解锁文章





