Pytorch 2.0, Cuda 11.8 and future 12.0 & 12.1 compatible

[johnnynunez]

I have been able to compile with CUDA 11.8 and CUDA 12.0 & 12.1 using PyTorch 1.13.1, PyTorch dev 2.0 and python 3.10.6(In local)

• Python <3.8 versions deleted (Deprecated 27 Jun 2023)
• Python 3.10.6 (added as default) (CircleCI images have 3.10.6 & 3.11)
• PyTorch <1.12 delete. Deprecated (https://pytorch.org/blog/deprecation-cuda-python-support/)
• Updated ubuntu-2204:2023.02.1 CPU
• Updated linux-cuda-11:2023.02.1 GPU https://discuss.circleci.com/t/cuda-11-8-gpu-cuda-image-any-plans/47240/3
• New image CircleCI CUDA 11.8, and they are working to CUDA 12. https://discuss.circleci.com/t/cuda-11-8-gpu-cuda-image-any-plans/47240/4
• Windows image 2022: windows-server-2022-gui:2022.10.1
• Updated test compatible with CUDA and PyTorch > 1.10
• Unified workflow GitHub and CircleCI python versions: Python 3.10.6
• Added CUDA variable on CircleCI: You now can select CUDA VERSION.
• Added PyTorch 2.0 and CUDA 11.8. 12.0/12.1 CircleCI (CircleCI working to release CUDA 12 images at the end of the month)
• Fixed tests for PyTorch 1.10+ and Python 3.10
• New versions and lint compatible with Python 3.10 [isort, black, flake8]
  
  

[johnnynunez]

@ppwwyyxx @wat3rBro @thiagocrepaldi Now, the changes is clearer.

[gianscarpe]

Up

[LeonidAlekseev]

up

[CanKorkut]

up

[elvizlai]

PTAL

[wadhah101]

Why this PR is not merged yet ? The current pytorch has some serious vulnerabilities

[wqh17101]

Please merge it in.

[johnnynunez]

Great work!
Are there any known backward incompatibilities (requiring reconfiguration or retraining)?
Are you perhaps also going to address any of the following?

• INSTALL.md: update table of prebuilt wheels
• docker/Dockerfile: update versions, maybe drop nvidia/cuda base image (as Pytorch does not use apt packages for CUDA and cuDNN anyway)
• setup.py treats torch as a setup-time dependency #4472 (i.e. fixing source distribution for platforms not covered by wheels, and adding torch as an overt dependency)

I'm not from Meta so I don't have access from links. But I updated the Install.md. @ppwwyyxx @facebook-github-bot merge it. It generates artifacts from macOS, Linux automatically.

Why do you need docker here? It should be in another PR

do you mean this?

[bertsky]

Why do you need docker here? It should be in another PR

Yes, maybe off-topic. But if it's not too much work, would be great to update as well (currently trying myself, will perhaps make a differential PR)

do you mean this?

Not quite. Since Pytorch (esp. the CUDA releases) may need a different index, and the dependency_links mechanism of setuptools has been dropped for good, I don't know how to cover this correctly. But in the very least, one would need to avoid the import torch statements further up in setup.py.

[johnnynunez]

for cuda releases you need extra index url

--extra-index-url https://download.pytorch.org/whl/cu118
onnx==1.14.0
torch==2.0.1+cu118; sys_platform != 'darwin'
torch==2.0.1; sys_platform == 'darwin'
onnxruntime==1.15.0; sys_platform == 'darwin' and platform_machine != 'arm64'
onnxruntime-silicon==1.13.1; sys_platform == 'darwin' and platform_machine == 'arm64'
onnxruntime-gpu==1.15.0; sys_platform != 'darwin'

[ZekunZh]

Up, @ppwwyyxx could you please approve and merge this PR ?

[janzd]

Could you please explain how to install using your wheels? Or if you have any installation instructions anywhere, could you send a link to it?

[johnnynunez]

Could you please explain how to install using your wheels? Or if you have any installation instructions anywhere, could you send a link to it?

download it and pip install name.whl

[danieltudosiu]

Any update on this?

[lijain]

I have been able to compile with CUDA 11.8 and CUDA 12.0 & 12.1 using PyTorch 1.13.1, PyTorch dev 2.0 and python 3.10.6(In local)

• Python <3.8 versions deleted (Deprecated 27 Jun 2023)
• Python 3.10.6 (added as default) (CircleCI images have 3.10.6 & 3.11)
• PyTorch <1.12 delete. Deprecated (https://pytorch.org/blog/deprecation-cuda-python-support/)
• Updated ubuntu-2204:2023.02.1 CPU
• Updated linux-cuda-11:2023.02.1 GPU https://discuss.circleci.com/t/cuda-11-8-gpu-cuda-image-any-plans/47240/3
• New image CircleCI CUDA 11.8, and they are working to CUDA 12. https://discuss.circleci.com/t/cuda-11-8-gpu-cuda-image-any-plans/47240/4
• Windows image 2022: windows-server-2022-gui:2022.10.1
• Updated test compatible with CUDA and PyTorch > 1.10
• Unified workflow GitHub and CircleCI python versions: Python 3.10.6
• Added CUDA variable on CircleCI: You now can select CUDA VERSION.
• Added PyTorch 2.0 and CUDA 11.8. 12.0/12.1 CircleCI (CircleCI working to release CUDA 12 images at the end of the month)
• Fixed tests for PyTorch 1.10+ and Python 3.10
• New versions and lint compatible with Python 3.10 [isort, black, flake8]
  
  

Excuse me, how do I download version 0.7?

[mac6894]

hi, I am using Ubuntu 22.04, CUDA 12.3, and would like to install detectron2. I did "python3 -m pip install 'git+https://github.com/johnnynunez/detectron2.git' ", but, I still get this runtime error:

"RuntimeError:
The detected CUDA version (12.3) mismatches the version that was used to compile
PyTorch (11.7). Please make sure to use the same CUDA versions.

  [end of output]"

Do you know how I can fix this? Thanks.

[johnnynunez]

hi, I am using Ubuntu 22.04, CUDA 12.3, and would like to install detectron2. I did "python3 -m pip install 'git+https://github.com/johnnynunez/detectron2.git' ", but, I still get this runtime error:

"RuntimeError: The detected CUDA version (12.3) mismatches the version that was used to compile PyTorch (11.7). Please make sure to use the same CUDA versions.

  [end of output]"

Do you know how I can fix this? Thanks.

First install pytorch:

pip3 install -U torch torchvision torchaudio