lib/libth_bert.so: undefined symbol

[pangsg]

I tried to run bert_example.py but a error raised

lib/libth_bert.so: undefined symbol: _ZN17fastertransformer13cublasAlgoMapC1ESsSs

I have build the FT with python bindings
Here is my build scripts:

cmake -DSM=80 -DCMAKE_BUILD_TYPE=Release -DBUILD_PYT=ON ..
make -j12

my env info:

GPU: A100
CUDA: 11.6
python:3.8.10
pytorch:1.12.1
fastertransformer: v5.0

[pangsg]

BTW, i couldn't use the provided docker image because of the policy of my company

[byshiue]

There are some issues in the environment settings. You can refer the recommended docker image to help debugging.

[pangsg]

Because of the rule of our company, I can't download the image from nvcr.io. Could plz provide some advice for debug?

[byshiue]

It is a undefined symbol issue due to environment issue, which depends on your environment settings and we are hard to provide suggestion.

You can try to check the cmake file and the make file generated by cmake, making sure they have linked successfully.

[pangsg]

Thanks for your advice. I have checked the built log, nothing failed in the log. Could u plz provide the requirements.txt?

[pangsg]

All Fixed.Thansk

[Siyuan011]

Hi, how do you fix it? I necountered the same problem.
Here is my config:
GPU: A100
CUDA: 11.3
python:3.8.5
pytorch:1.12.1

[pangsg]

Hi, how do you fix it? I necountered the same problem. Here is my config: GPU: A100 CUDA: 11.3 python:3.8.5 pytorch:1.12.1

do u follow the xx_guide.md in docs?

[Siyuan011]

yes of course. I followed gpt_guide

[pangsg]

yes of course. I followed gpt_guide

hmm, i followed the guide and fixed the bug. The git is necessary