Pytorch CUDNN issues

AI & Data Science Deep Learning (Training & Inference) cuDNN
, , ,
1

I get a error when i try to use a nn.ConvTranspose2d layer in pytorch in my model while training it.

Error:
Error: /usr/local/cuda/lib64/libcudnn_cnn_train.so.8: undefined symbol: _ZN5cudnn3cnn34layerNormFwd_execute_internal_implERKNS_7backend11VariantPackEP11CUstream_stRNS0_18LayerNormFwdParamsERKNS1_20NormForwardOperationEmb, version libcudnn_cnn_infer.so.8

cudnn is correctly installed. i tested it using tensorflow but i can’t convert my model to tf

2

Hi @jeremy.naumann2007 ,
Checking with Engineering team.

4

Hi @jeremy.naumann2007 ,
Looks like there is mixing of different cuDNN versions happening (e.g. the one shipped in the PyTorch wheels and another one from the system lib path). can you please confirm that?