Skip to content

Conversation

peterjc123
Copy link
Contributor

No description provided.

@codecov-io
Copy link

codecov-io commented Aug 14, 2019

Codecov Report

Merging #1241 into master will decrease coverage by 0.06%.
The diff coverage is n/a.

Impacted file tree graph

@@            Coverage Diff             @@
##           master    #1241      +/-   ##
==========================================
- Coverage   65.84%   65.77%   -0.07%     
==========================================
  Files          75       75              
  Lines        5785     5785              
  Branches      885      885              
==========================================
- Hits         3809     3805       -4     
- Misses       1710     1713       +3     
- Partials      266      267       +1
Impacted Files Coverage Δ
torchvision/transforms/functional.py 70.11% <0%> (-1.15%) ⬇️
torchvision/transforms/transforms.py 80.98% <0%> (ø) ⬆️

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 6de158c...aad9b82. Read the comment docs.

@peterjc123
Copy link
Contributor Author

peterjc123 commented Aug 14, 2019

@ezyang The yaml file for Azure Pipelines is added as packaging/windows/azure-pipelines-ci.yaml.

echo "CUDA version should be M.m with no dot, e.g. '8.0' or 'cpu'"
exit 1
fi

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Heads up: this script is no longer used by the Linux/OS X binary builds (I didn't delete it because Windows still needs it), instead packaging/build_conda.sh is what is used. It would pretty awesome if we could also unify these packaging scripts, but I only really want to do it if we can keep the packaging on the shorter side. In the meantime, you can rename / simplify this script so that it is Windows only.

- NVCC_FLAGS
features:
{{ CONDA_CPUONLY_FEATURE }}
{{ environ.get('CONDA_CPUONLY_FEATURE') }}
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nice catch, I wonder how this ever worked lol

Copy link
Contributor

@ezyang ezyang left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I gave you commit bits so you can merge these changes directly

@fmassa
Copy link
Member

fmassa commented Sep 6, 2019

@peterjc123 what's the status of this PR?

@peterjc123
Copy link
Contributor Author

It runs fine for building binaries, but for nightlies, it requires our own machines, which cannot be satisfied.

@fmassa
Copy link
Member

fmassa commented Sep 9, 2019

@peterjc123 so this is a superset of the Windows improvements you did in #1298?

@peterjc123
Copy link
Contributor Author

yes, you could say that. But it was actually a different direction, which tries to do releases using Azure Pipelines, not CircleCI.

@fmassa
Copy link
Member

fmassa commented Sep 9, 2019

@peterjc123 from the discussion in pytorch/pytorch#24859 (comment), what do you think about updating this patch to instead run everything on CircleCI?

cc @ezyang

@peterjc123
Copy link
Contributor Author

The problem is the limited parallelism (only 10x). @ezyang Have you talked about it with a related person?

@ezyang
Copy link
Contributor

ezyang commented Sep 9, 2019

@peterjc123 Wait, I thought you were referring to Azure parallelism when you said we only got 10x parallelism. I can file a ticket with CircleCI to check what our Windows parallelism is, but I'm pretty sure we get more than 10x.

@ezyang
Copy link
Contributor

ezyang commented Sep 9, 2019

You'll have to actually run the builds on PRs on these repositories, though, to get our limits.

@peterjc123
Copy link
Contributor Author

peterjc123 commented Sep 10, 2019

@ezyang Well, I misunderstood what @fmassa said. I was thinking that he wanted me to continue doing releases using Azure Pipelines. But actually, he needs CircleCI instead. I think it's feasible to add GPU CI for Windows. However, for releases, we may need to install some old versions of CUDA that is unsupported in CircleCI. (Windows Server 2019 is currently the only supported option for Windows on CircleCI and only CUDA 10.1 supports Windows Server 2019.) What's more, old CUDA leads to old VS and I haven't managed to install them in a Windows container. (The technology CircleCI based on)

@ezyang
Copy link
Contributor

ezyang commented Sep 10, 2019

@soumith Do you know what our planned support window for CUDA 9.2 will be? My guess is we'd keep supporting it until CUDA 11 comes out...

@peterjc123
Copy link
Contributor Author

@ezyang @fmassa The CPU and CUDA nightly build jobs are now working on Microsoft hosted agents. Do you want to enable them?

Copy link
Contributor Author

@peterjc123 peterjc123 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@ezyang Could you please upload these binaries to ossci-windows?

curl -k -L https://www.dropbox.com/s/z5b7ryz0zrimntl/cuda_9.0.176_windows.7z?dl=1 --output "%SRC_DIR%\temp_build\cuda_9.0.176_windows.7z"
:cuda92
if not exist "%SRC_DIR%\temp_build\cuda_9.2.148_win10.exe" (
curl -k -L https://developer.nvidia.com/compute/cuda/9.2/Prod2/local_installers2/cuda_9.2.148_win10 --output "%SRC_DIR%\temp_build\cuda_9.2.148_win10.exe"
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

+1

IF NOT EXIST "%SRC_DIR%\temp_build\cudnn-9.0-windows7-x64-v7.zip" (
curl -k -L https://www.dropbox.com/s/6p0xyqh472nu8m1/cudnn-9.0-windows7-x64-v7.zip?dl=1 --output "%SRC_DIR%\temp_build\cudnn-9.0-windows7-x64-v7.zip"
if not exist "%SRC_DIR%\temp_build\cudnn-9.2-windows10-x64-v7.2.1.38.zip" (
curl -k -L https://downloads.sourceforge.net/project/cuda-dnn/7/CUDA-9.2/cudnn-9.2-windows10-x64-v7.2.1.38.zip --output "%SRC_DIR%\temp_build\cudnn-9.2-windows10-x64-v7.2.1.38.zip"
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

+1

IF NOT EXIST "%SRC_DIR%\temp_build\cuda_9.1.85_windows.7z" (
curl -k -L https://www.dropbox.com/s/7a4sbq0dln6v7t2/cuda_9.1.85_windows.7z?dl=1 --output "%SRC_DIR%\temp_build\cuda_9.1.85_windows.7z"
if not exist "%SRC_DIR%\temp_build\cuda_10.0.130_411.31_win10.exe" (
curl -k -L https://developer.nvidia.com/compute/cuda/10.0/Prod/local_installers/cuda_10.0.130_411.31_win10 --output "%SRC_DIR%\temp_build\cuda_10.0.130_411.31_win10.exe"
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

+1

IF NOT EXIST "%SRC_DIR%\temp_build\cudnn-9.1-windows7-x64-v7.zip" (
curl -k -L https://www.dropbox.com/s/e0prhgsrbyfi4ov/cudnn-9.1-windows7-x64-v7.zip?dl=1 --output "%SRC_DIR%\temp_build\cudnn-9.1-windows7-x64-v7.zip"
if not exist "%SRC_DIR%\temp_build\cudnn-10.0-windows10-x64-v7.4.1.5.zip" (
curl -k -L https://www.dropbox.com/s/9v1z9rmbjw9mhx2/cudnn-10.0-windows10-x64-v7.4.1.5.zip?dl=1 --output "%SRC_DIR%\temp_build\cudnn-10.0-windows10-x64-v7.4.1.5.zip"
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

+1

@peterjc123
Copy link
Contributor Author

Future work: We can use Windows container to save time (~14 min to install VS and CUDA).

@ezyang
Copy link
Contributor

ezyang commented Sep 16, 2019

Could you please upload these binaries to ossci-windows?

These are nightly binaries right? We should just setup automatic upload. Does Azure have a way to add private environment variables? If so, I can give you S3 creds to the nightlies bucket.

Copy link
Contributor

@ezyang ezyang left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sure, go for it!

@peterjc123
Copy link
Contributor Author

peterjc123 commented Sep 16, 2019

Could you please upload these binaries to ossci-windows?

These are nightly binaries right? We should just setup automatic upload. Does Azure have a way to add private environment variables? If so, I can give you S3 creds to the nightlies bucket.

No, these are the installers of CUDA and cuDNN. According to the packer-windows scripts, they should have of the URL of https://s3.amazonaws.com/ossci-windows/win2016/******. I'll wait until these binaries are uploaded and then change the URLs in this PR.

@peterjc123
Copy link
Contributor Author

@ezyang Looks like I've uploaded these wheels by mistake, could you please remove them in https://download.pytorch.org/whl/nightly/cpu?
torchvision-0.5.0a0%2B6de158c-cp35-cp35m-win_amd64.whl
torchvision-0.5.0a0%2B6de158c-cp36-cp36m-win_amd64.whl
torchvision-0.5.0a0%2B6de158c-cp37-cp37m-win_amd64.whl

@ezyang
Copy link
Contributor

ezyang commented Sep 17, 2019

Looks like I've uploaded these wheels by mistake, could you please remove them in https://download.pytorch.org/whl/nightly/cpu?

Done

@ezyang
Copy link
Contributor

ezyang commented Sep 17, 2019

No, these are the installers of CUDA and cuDNN. According to the packer-windows scripts, they should have of the URL of https://s3.amazonaws.com/ossci-windows/win2016/******. I'll wait until these binaries are uploaded and then change the URLs in this PR.

I misunderstood your comment :) Working on it.

@peterjc123 peterjc123 merged commit 490966f into pytorch:master Sep 18, 2019
@peterjc123 peterjc123 deleted the win_fix branch September 18, 2019 06:29
@peterjc123
Copy link
Contributor Author

@fmassa @ezyang Nightly builds are now scheduled for Windows every night.

@fmassa
Copy link
Member

fmassa commented Sep 18, 2019

Awesome, thanks a lot @peterjc123 !

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants