Change CI cuda versions to 10.2

[datumbox]

Updates the CI to use cuda 10.2 instead of 10.1 so that we can start using fresh PyTorch core nightlies.

[NicolasHug]

Thanks for the PR!

Let's see what the CI says. We might need to also change stuff like the use of image_name: "pytorch/manylinux-cuda101" in the yaml file, as well as other references to 101 in regenerate.py, in particular:

                if device_type == 'gpu':
                    if python_version != "3.8":
                        job['filters'] = gen_filter_branch_tree('master', 'nightly')
                    job['cu_version'] = 'cu101'

[datumbox]

Thanks I'm still looking for more places I have to change. If you see others please let me know.

[datumbox]

It seems that the latest PyTorch core on linux has a new restriction on how you index things across devices:

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking arugment for argument indices in method wrapper_Tensor_index_Tensor)

The problem appears only on Linux GPU and not on Windows:
https://app.circleci.com/pipelines/github/pytorch/vision/8241/workflows/6af4ce74-e157-4ca6-9eca-64ed7b9989ee/jobs/590609/tests#failed-test-0

I propose to merge this now and fix ASAP the issues on master on a separate PR.

[fmassa]

Accepting to unblock

[ngimel]

To close the loop on this, indexing from python is not affected, because python_variable_indexing copies all the indices to self device before dispatching to index. However, having indices on the different device in c++ has been broken by pytorch/pytorch#56872. cc @wenleix, @ezyang

[wenleix]

@ngimel Thanks for reporting this. Will figure out a fix to this.

Update: index leverage TensorIterator and should be opt-out for automatic device check:

https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/TensorAdvancedIndexing.cpp#L296

make_index_iterator:

https://github.com/pytorch/pytorch/blob/dc8bc6ba4bbd61d21de4ffbccf5b79d22ff31a23/aten/src/ATen/native/TensorAdvancedIndexing.cpp#L266-L277