CE模型: sequence_tagging_for_ner 执行模型出错

[guochaorong]

http://18.222.34.7:8080/viewLog.html?buildId=881&buildTypeId=Paddle_ContinuousEvaluation&tab=buildLog 这是错误日志

从11102 开始这个错误出现

11102 代码
#11102

复现方法：
登录到CE web
http://18.222.34.7:8080/
点击

进入后， 点击右上角run右边， 选择 run的pr id

[panyx0718]

@tensor-tang 如果短期内修不好就先revert?

[tensor-tang]

我没意见，这个是mkldnn的新feature。

并且我还在定位问题，他并没有改变原有逻辑，只是加了use_mkldnn=false这些attr。

[tensor-tang]

@mozga-intel , this should be caused by your commits in #11102

How to reproduce on your local machine:

1. build paddle with WITH_GPU=ON and install
2. git clone https://github.com/PaddlePaddle/paddle-ce-latest-kpis.git && cd
   sequence_tagging_for_ner
3. remove from continuous_evaluation import * at train.py
4. then run.xsh

As a hot fix, I would revert this feature at this moment.

[mozga-intel]

@tensor-tang Thank you, I will reproduce it as soon as possible.