GCP Base Image Wrong CUDA Version

[PetrochukM]

🐛 Bug

Working through the suggested PyTorch GCP base image, the wrong version of CUDA was installed.

To Reproduce

Steps to reproduce the behavior:

1. On "New Deep Learning VM deployment" select "Intel Optimized Base 9.2"
2. Deploy and SSH into the machine
3. nvidia-smi shows CUDA Version is at 10.0

Expected behavior

The CUDA Version is correct after setting up a GCP machine.

[PetrochukM]

Let me know if this is the wrong form to report the bug...

[colesbury]

@PetrochukM is this an issue with Google Cloud Platform? Or an issue with PyTorch documentation?

Where is the "suggested PyTorch GCP base image"?

[PetrochukM]

The issue is with GCP; however, it's a solution suggested on https://pytorch.org/ under "CLOUD PARTNERS".

[colesbury]

@soumith can you route this to a contact at GCP?

[soumith]

@b0noI this seems like the wrong binary got installed. can you take a look?

[mikehcheng]

I work with @b0noI, the image is working as intended, but the user is likely noticing the behavior where nvidia-smi shows the driver version instead of the CUDA version. See https://stackoverflow.com/questions/53422407/different-cuda-versions-shown-by-nvcc-and-nvidia-smi for a more detailed explanation.

That being said, for PyTorch we recommend selecting the framework "PyTorch 1.0 + fast.ai 1.0 (CUDA 10.0)". The other images do not have PyTorch installed.