Hap-Zhang

Hi,all i want to load a Hclr.fst by the code "hcl_fst = fst::StdFst::Read(hcl_fst_rxfilename)", and get a error as "ERROR: GenericRegister::GetEntry: lookup failed in shared object: olabel_lookahead-fst.so", would you like to...

Yes,i config openfst-1.6.1 with option --enable-lookahead-fsts and build&install recently.I want to try lookahead graph decoding when online decoding,and unluckily i get this error.

ok,i'll have a try on the OpenFST forums firstly,thank you.

@Malkovsky Yeah,i see. I really export paths like:`export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$KALDI_ROOT/tools/openfst/lib/fst` However, the error is:"lookup failed in shared object: olabel_lookahead-fst.so", and not "olabel_lookahead-fst.so: cannot open shared object file: No such file or...

@feifeibear 好的，谢谢大佬。 我看代码里这块单独处理了pooler层，是有什么讲究吗？  我基于huggingface预训练模型进行finetuning，出来的模型用turbo作为后端运行，会报下面的错误，后来调试发现模型中其实没有pooler这一层的，TurboTransformer的代码里为什么会特地加入pooler呢？ 

我这边用的预训练模型是bert-base-chinese

嗯，的确，我用的是BertForTokenClassification，这个里面没有用到pooler，那是不是意味着我用的话，需要改下turbotransformers底层的代码呢？

请问中文大概多少个字能体现出效果呢，我可以再测试下 onnxrt是有些加速的，虽然我不知道为什么onnxrt这边Torch的时长更长了。。。。 

我这边把timeline打出来了，您可以帮忙看看哪块有比较大的嫌疑吗？ 

好的，OMP线程数目默认是机器本身CPU个数吗？