Torch==1.6.0 on Google cloud with CUDA

Sorry if this is a newb question. I don’t have a lot of experience with GPUs…

I created a GCE instance like this:

gcloud compute instances create my-instance --image-family=pytorch-latest-gpu ...

The instance that spins up has torch 1.4 instead of the latest one. With this version, torch.cuda.is_available() returns True.

After doing pip install -U torch to get the latest version, torch.cuda.is_available() returns False.

What is the best way to get a working torch==1.6.0 instance on GCE?

I would assume this command should install the latest stable PyTorch version with CUDA10.2.
Which GPUs are you using in the node? Maybe the compute capability is not shipped anymore in the binaries?

I’m using a V100 GPU.

I’m guessing that torch 1.4 and 1.6 use different CUDA versions so that the previously installed drivers don’t work with torch 1.6.

It is a bummer that Google isn’t updating the pytorch-latest-gpu image. :frowning:

Could you post the install log from the notebook when you are trying to install the latest stable PyTorch version?