The Release 0.13 is not using the correct torch core version

[YosuaMichael]

🐛 Describe the bug

Currently the CI on branch release/0.13 is failing (I create dummy PR #6092 to track this).
We fail on type_check and this does not happen on main branch.
After looking at the error: https://app.circleci.com/pipelines/github/pytorch/vision/17906/workflows/8f008f5a-13b2-4207-a378-51ae705b99b7/jobs/1450718
especially on the installation of pytorch, we get:

#!/bin/bash -eo pipefail
pip install  --user  --progress-bar=off --pre torch --extra-index-url https://download.pytorch.org/whl/test/cpu/torch_test.html

Looking in indexes: https://pypi.org/simple, https://download.pytorch.org/whl/test/cpu/torch_test.html
Collecting torch
  Downloading torch-1.11.0-cp37-cp37m-manylinux1_x86_64.whl (750.6 MB)

We should use the RC branch 1.12 from pytorch core but instead we get version 1.11.

After digging a bit deeper, if we search "torch-" on https://download.pytorch.org/whl/test/cpu/torch_test.html currently I can only see 4 versions and the latest one is indeed 1.11. I think there is a problem here.

cc @malfet @seemethere @datumbox @atalman

[pmeier]

Maybe I'm missing something here, but there are plenty of 1.12.0 binaries in the list:

torch-1.12.0%2Bcpu-cp310-cp310-linux_x86_64.whl
torch-1.12.0%2Bcpu-cp310-cp310-win_amd64.whl
torch-1.12.0%2Bcpu-cp37-cp37m-linux_x86_64.whl
torch-1.12.0%2Bcpu-cp37-cp37m-win_amd64.whl
torch-1.12.0%2Bcpu-cp38-cp38-linux_x86_64.whl
torch-1.12.0%2Bcpu-cp38-cp38-win_amd64.whl
torch-1.12.0%2Bcpu-cp39-cp39-linux_x86_64.whl
torch-1.12.0%2Bcpu-cp39-cp39-win_amd64.whl
torch-1.12.0-cp310-none-macosx_10_9_x86_64.whl
torch-1.12.0-cp310-none-macosx_11_0_arm64.whl
torch-1.12.0-cp37-none-macosx_10_9_x86_64.whl
torch-1.12.0-cp37-none-macosx_11_0_arm64.whl
torch-1.12.0-cp38-none-macosx_10_9_x86_64.whl
torch-1.12.0-cp38-none-macosx_11_0_arm64.whl
torch-1.12.0-cp39-none-macosx_10_9_x86_64.whl
torch-1.12.0-cp39-none-macosx_11_0_arm64.whl
torch-1.12.0.dev20220520%2Bcpu-cp310-cp310-linux_x86_64.whl
torch-1.12.0.dev20220520%2Bcpu-cp310-cp310-win_amd64.whl
torch-1.12.0.dev20220520%2Bcpu-cp37-cp37m-linux_x86_64.whl
torch-1.12.0.dev20220520%2Bcpu-cp37-cp37m-win_amd64.whl
torch-1.12.0.dev20220520%2Bcpu-cp38-cp38-linux_x86_64.whl
torch-1.12.0.dev20220520%2Bcpu-cp38-cp38-win_amd64.whl
torch-1.12.0.dev20220520%2Bcpu-cp39-cp39-linux_x86_64.whl
torch-1.12.0.dev20220520%2Bcpu-cp39-cp39-win_amd64.whl
torch-1.12.0.dev20220520-cp310-none-macosx_10_9_x86_64.whl
torch-1.12.0.dev20220520-cp310-none-macosx_11_0_arm64.whl
torch-1.12.0.dev20220520-cp37-none-macosx_10_9_x86_64.whl
torch-1.12.0.dev20220520-cp37-none-macosx_11_0_arm64.whl
torch-1.12.0.dev20220520-cp38-none-macosx_10_9_x86_64.whl
torch-1.12.0.dev20220520-cp38-none-macosx_11_0_arm64.whl
torch-1.12.0.dev20220520-cp39-none-macosx_10_9_x86_64.whl
torch-1.12.0.dev20220520-cp39-none-macosx_11_0_arm64.whl

[pmeier]

The installation command is wrong. It should be either

-f https://download.pytorch.org/whl/test/cpu/torch_test.html

or

--extra-index-url https://download.pytorch.org/whl/test/cpu

The former only works if we install PyTorch binaries exclusively, while the second can also pull stuff from PyPI if it doesn't find anything in the extra index.

[pmeier]

This was introduced in #6080. The only change that is needed is

diff --git a/.circleci/config.yml b/.circleci/config.yml
index d0e3d881e..d1120f45c 100644
--- a/.circleci/config.yml
+++ b/.circleci/config.yml
@@ -142,7 +142,7 @@ commands:
         default: true
     steps:
       - pip_install:
-          args: --pre torch --extra-index-url https://download.pytorch.org/whl/test/cpu/torch_test.html
+          args: --pre torch --extra-index-url https://download.pytorch.org/whl/test/cpu
           descr: Install PyTorch from nightly releases
       - pip_install:
           args: --no-build-isolation <<# parameters.editable >> --editable <</ parameters.editable >> .
diff --git a/.circleci/config.yml.in b/.circleci/config.yml.in
index eb0d2e08c..cf0077a49 100644
--- a/.circleci/config.yml.in
+++ b/.circleci/config.yml.in
@@ -142,7 +142,7 @@ commands:
         default: true
     steps:
       - pip_install:
-          args: --pre torch --extra-index-url https://download.pytorch.org/whl/test/cpu/torch_test.html
+          args: --pre torch --extra-index-url https://download.pytorch.org/whl/test/cpu
           descr: Install PyTorch from nightly releases
       - pip_install:
           args: --no-build-isolation <<# parameters.editable >> --editable <</ parameters.editable >> .

@atalman Do you want me to send a PR?

[YosuaMichael]

Maybe I'm missing something here, but there are plenty of 1.12.0 binaries in the list:

Ah you are right @pmeier , seems like I missed it because I searched for /torch- instead. My bad!

[atalman]

@pmeier I will send a PR now and do another branch cut

[NicolasHug]

@atalman @malfet @pmeier

In https://github.com/pytorch/vision/pull/5672/files, we switched from -f to --extra-index-url. See in particular this discussion between @pmeier and I https://github.com/pytorch/vision/pull/5672/files#r834412238.

It's clear that at least for the release branch, we should be using -f. It seems to me that we should also revert to -f in main, as it would at least avoid future mistakes like #6080, where I based my changes off the diff from https://github.com/pytorch/vision/pull/4496/files, and missed the subtle change of -f.

Could you please advise on this?

[pmeier]

AFAIK, we no longer advertise the -f syntax. It is probably only still working for BC.

It's clear that at least for the release branch, we should be using -f.

Is your reason here to avoid pulling any binary other than the test binaries? If that is the case, we can simply switch to --index-url instead of --extra-index-url. This allows us to keep the documented index syntax, while also being sure not to pull anything that is not hosted by us.

[atalman]

@NicolasHug @YosuaMichael, I think this issue is resolved, Can we close it ?

[NicolasHug]

The original issue is resolved, but if you don't mind I'd love your thoughts on #6099 (comment) before closing