Upgrading PyTorch to v1.8.0 and torchvision to v0.9.0

[seemethere]

Hi folks,

We released a new PyTorch version v1.8.0 and torchvision v0.9.0.

The wheel links are:

PyTorch

CUDA 10.1

download.pytorch.org/whl/cu101/torch-1.8.0%2Bcu101-cp36-cp36m-linux_x86_64.whl

CUDA 10.2

download.pytorch.org/whl/cu102/torch-1.8.0-cp36-cp36m-linux_x86_64.whl

TorchVision

CUDA 10.1

download.pytorch.org/whl/cu101/torchvision-0.9.0%2Bcu101-cp36-cp36m-linux_x86_64.whl

CUDA 10.2

download.pytorch.org/whl/cu102/torchvision-0.9.0-cp36-cp36m-linux_x86_64.whl

Other notes

If you require wheels for Python 3.7 / 3.8, you can find the links here: download.pytorch.org/whl/torch_stable.html

We're looking to having it updated in Colab.

Thanks very much.

cc'ing @colaboratory-team

Similar to #1454

[colaboratory-team]

Previous update was #1676.
@seemethere colab is now on Python 3.7 and CUDA 11.0; I'm not seeing 1.8.0 wheels under cu110/ at https://download.pytorch.org/whl/torch_stable.html (only under cu111/, presumably meaning CUDA 11.1). Can you release 11.0 wheels?

[seemethere]

Previous update was #1676.
@seemethere colab is now on Python 3.7 and CUDA 11.0; I'm not seeing 1.8.0 wheels under cu110/ at download.pytorch.org/whl/torch_stable.html (only under cu111/, presumably meaning CUDA 11.1). Can you release 11.0 wheels?

Unfortunately we've dropped CUDA 11.0 in favor of CUDA 11.1 and have no plans to release binaries for CUDA 11.0

[colaboratory-team]

Ok, we can use your 10.1 wheels then:
https://download.pytorch.org/whl/cu101/torch-1.8.0%2Bcu101-cp37-cp37m-linux_x86_64.whl
https://download.pytorch.org/whl/cu101/torchvision-0.9.0%2Bcu101-cp36-cp36m-linux_x86_64.whl

Tracked internally by b/181907109.

[mthrok]

Hi @colaboratory-team

Could you consider adding torchaudio preinstalled alongside of PyTorch/torchvision?

[craigcitro]

@mthrok In general, we tend to be conservative with the packages we install, as each package adds complexity on our side (keeping versions compatible, install time, disk usage, etc).

For something like torch or torchvision, where it's important that the installed .whl match the installed CUDA version, it makes sense for us to own that, as we can keep it in sync with changes on our side.

For torchaudio, though, there's less of a compatibility question (no GPU-using code AFAICT), and the install is fast (!pip install torchaudio took 5 seconds in a fresh runtime). In that case, "explicit is better than implicit" wins -- users can install it, and pin a version that works for them, making it less likely that we'll silently break them with an update in the future.

[mthrok]

@craigcitro I see. That totally makes sense and that decision model is easy to get. Thanks for the explanation.

[craigcitro]

Pytorch and torchvision should be updated.