[CI] Torch 1.7 update to mainline

[masahi]

Closes #6594

Reopening #6811 to reflect the change in CI image. Tests in the staging branch have passed with the new image https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/ci-docker-staging/34/pipeline

Now, ci-gpu should be running v0.72, the new image with torch 1.7

$ docker run -it tlcpack/ci-gpu:v0.72
root@d64e8b20bae4:/# python3
Python 3.6.12 (default, Aug 18 2020, 02:08:22) 
[GCC 5.4.0 20160609] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.__version__
'1.7.0'
>>> import torchvision
>>> torchvision.__version__
'0.8.1'

[masahi]

I've hit a flaky test introduced in #6750
https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-6828/1/pipeline/252

@jcf94 @merrymercy @tqchen

[jcf94]

I've hit a flaky test introduced in #6750
https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-6828/1/pipeline/252

@jcf94 @merrymercy @tqchen

Thanks! I'm also checking this part.

[masahi]

@jcf94 how is it looking? Shall I disable those tests for now?

[jcf94]

@jcf94 how is it looking? Shall I disable those tests for now?

Really sorry for this, seems this bug only occured in the CI runtime.
You can currently disable it for not blocking other PRs.

[masahi]

yeah no problem, this seems i386 specific, I imagine it would be hard to repro.

[jcf94]

yeah no problem, this seems i386 specific, I imagine it would be hard to repro.

Thanks.

[masahi]

tvm-ci/branch (ci-docker-staging) failed with no space left on device error, but since the tests on ci-docker-staging have already passed in https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/ci-docker-staging/34/pipeline/ and the image is updated, it doesn't matter (sending a PR from ci-docker-staging was my mistake, causing unnecessary CI tests).

@tqchen @zhiics @kevinthesun Ready to land PyTorch 1.7 support !!