Hi @fancyerii ,

Thanks for reporting.Its known issue.It seems this might be due to duplicate registry AFAIK. Though its logging an error it seems the GPU works fine.In your logs also the GPUs also detected and it will not affecting the execution. Could you confirm whether there is a problem with execution of your code?

Same issue was discussed in #62002, #62075.May please refer for more responses there. Thanks!

CC: @learning-to-play

@fancyerii , Just to verify please cross check whether cudnn_version: 8 & cuda_version 12.2 are there or not using the code below.

from tensorflow.python.platform import build_info

print("cudnn_version",build_info.build_info['cudnn_version'])
print("cuda_version",build_info.build_info['cuda_version'])

@SuryanarayanaY , I'm having the same problem as the OP on latest-gpu, just run your recommended cross check and am getting:

>>> print("cudnn_version",build_info.build_info['cudnn_version'])
cudnn_version 8
>>> print("cuda_version",build_info.build_info['cuda_version'])
cuda_version 12.2

So you're saying I just should ignore the "unable to register..." then? They are the only reason I at all went through the trouble of attempting to use Docker images... I'll revert back to my native install then. Thanks.

Note: lines starting with "E" tend to be interpreted as "errors", I guess, i.e. possibly blocking problems. I guess this makes this issue quite a serious one.

In fact, there is one more difference for me between native and Docker. With native, I am also getting:

W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT

And that seems to be fixed by neither a regular nor a local pip install of tensorrt and friends. So I guess the Docker image still has an added value for me. Unless there is an obvious fix for the local TensorRT problem (I am installing TensorRT according to NVIDIA's instructions as far as I know)?

same problem here :(