Closed wtrainor closed 3 years ago
There is no issue with the Dockerfile. You need to have the CUDA driver installed on the host machine and then run docker using the --runtime=nvidia
argument to tell docker to mount the host drivers.
The Dockerfile is just for convenience, if you have all dependencies locally installed you can run it directly on the host machine.
I ran : docker run -it pytorch/pytorch:1.1.0-cuda10.0-cudnn7.5-runtime /bin/bash
And : echo $CUDA_HOME
with an empty result.
So pytorch cannot find the CUDA version. I believe it's an issue with the Dockerfile. Can you advise? Can I run without the docker?