-
Notifications
You must be signed in to change notification settings - Fork 14
Open
Description
Description:
I'm encountering an error when attempting to run the sample Policy RL script. It seems to be related to utilizing CUDA on a GPU. The error message indicates a failure to initialize the DNN library.
UnfilteredStackTrace: jaxlib.xla_extension.XlaRuntimeError: FAILED_PRECONDITION:
DNN library initialization failed. Look at the errors above for more details.
Environment:
OS: Ubuntu 22.04
CUDA Version: 11.8
JAX Version: 0.4.13
JAXlib Version: 0.4.13+cuda11.cudnn86
Python Version: Python 3.8.19
Steps to set up
- I created a conda environment and installed the Cuda 11.8 toolkit using
conda install cuda -c nvidia/label/cuda-11.8.0 - Then I install PyTorch using
pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu118 - Then I install jax with GPU support for Cuda 11 using
pip install --upgrade "jax[cuda11_pip]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html - Then, I install the repo dependencies
pip install -r requirements.txt
Attempts to narrow issue
- I verified that the GPU is accessible using
torch.cuda.is_available(): returns True. - I verified that jax can detect the the GPU using 'import jax' and 'jax.devices': returns [gpu(id=0)]
Stack Trace
...
Metadata
Metadata
Assignees
Labels
No labels




