Closed mindbound closed 3 months ago
hmaarrfk
commented
3 months ago Can you try again in a new environment with:
conda create --name tf_issue_380 python=3.11 tensorflow --channel conda-forge --override-channelsi see some things from nvidia, and pytorch channels which can create incompatibility
mindbound
commented
3 months ago I can import it in the fresh environment, although I am getting some warnings:
2024-04-13 16:06:58.807341: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:9261] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
2024-04-13 16:06:58.807509: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:607] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
2024-04-13 16:06:58.811954: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1515] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registeredWhat might be the possible culprits in my existing environment and what would be the best course of action? Should I create separate environments for PyTorch and TF?
hmaarrfk
commented
3 months ago What might be the possible culprits in my existing environment
we don't support environments from mixed channels. So I can't really help you answer you questions. the "main", "conda-forge", "nvidia", "pytorch" channels all have different ways of organizing libraries and require different c level packages.
what would be the best course of action? Should I create separate environments for PyTorch and TF?
Thats up to you.
TF and Pytorch are ML Training libraries. Do you really need to train a model accross both?
hmaarrfk
commented
3 months ago My suggestion is to stick to the conda-forge channel and help package what you need here. Though I am biased.
mindbound
commented
3 months ago Do you really need to train a model accross both?
Unfortunately yes.
My suggestion is to stick to the conda-forge channel and help package what you need here.
Thanks! I'll try rebuilding the environment using coda-forge only.
hmaarrfk
commented
3 months ago Unfortunately yes.
Basically, each project will declare its dependencies.
Pytorch, nvidia, tensorflow upstream instructions declare "default analconda"
conda-forge doesn't.
So you could also use default anaconda and not conda-forge.
Good luck!
Solution to issue cannot be found in the documentation.
Issue
When importing TensorFlow as
import tensorflow as tf, the following error occurs:All packages are up to date.
Installed packages
Environment info