fix mask and position bug for batch generation

#35
by qingsonglv - opened
Knowledge Engineering Group (KEG) & Data Mining at Tsinghua University org
No description provided.

这里确实是有问题的,batch的时候,官方没有修复这个bug。你的修改了之后训练起来还有问题吗?

Knowledge Engineering Group (KEG) & Data Mining at Tsinghua University org
edited Apr 11, 2023

训练是不受影响的,因为训练没有left pad,只是inference有影响。batch size=1的chat函数应该也没有影响,只有不定长batch generation的时候有影响。

Knowledge Engineering Group (KEG) & Data Mining at Tsinghua University org

看起来是我这边测试的问题,其实是没有bug的……

qingsonglv changed pull request status to closed

Sign up or log in to comment