[Training] pip install onnxruntime.training still in 1.15.1

[elephantpanda]

Describe the issue

Do: pip install onnxruntime.training in Windows It is still on 1.15.1 whereas it should be 1.17.1

To reproduce

pip install onnxruntime.training

Urgency

No response

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.17.1

PyTorch Version

1.13.1

Execution Provider

CUDA

Execution Provider Library Version

No response

[AdamLouly]

What are you trying to use onnxruntime training for? if its on device training please follow instructions here: https://onnxruntime.ai/getting-started

if its for large model Training using ORTModule, we advise to use WSL if you're using windows.

[elephantpanda]

Thanks for asking. I am trying to create training artifacts. Your link, I believe, does not automatically install the onnxruntime.training module

Unfortunately some layers such as ConvTransposeGrad have not been implemented in 1.15.1 only in the later version.

Hence why I'm trying to install onnxruntime.traning 1.17.1. But for some reason this is not available yet. (I have been waiting patiently for about 6 months hoping it would be updated :) )

I would like to create the training artifacts so I can do train-on-device using C#.

The pip install onnxruntime-training says it is version 1.17.1 but I tried it on Windows and it is only installing 1.15.1:

ERROR: Could not find a version that satisfies the requirement onnxruntime-training==1.17.1 (from versions: 1.15.0, 1.15.1)
ERROR: No matching distribution found for onnxruntime-training==1.17.1

[AdamLouly]

You can follow the instructions here to build from source for on device training. https://onnxruntime.ai/docs/build/training.html

 git clone --recursive https://github.com/Microsoft/onnxruntime.git
 cd onnxruntime
 ./build.sh --config RelWithDebInfo --build_shared_lib --parallel --enable_training_apis

@askhade do you know why we're not publishing the recent stable version for windows?

[baijumeswani]

Thanks for asking. I am trying to create training artifacts. Your link, I believe, does not automatically install the onnxruntime.training module

Unfortunately some layers such as ConvTransposeGrad have not been implemented in 1.15.1 only in the later version.

Hence why I'm trying to install onnxruntime.traning 1.17.1. But for some reason this is not available yet. (I have been waiting patiently for about 6 months hoping it would be updated :) )

I would like to create the training artifacts so I can do train-on-device using C#.

The pip install onnxruntime-training says it is version 1.17.1 but I tried it on Windows and it is only installing 1.15.1:

ERROR: Could not find a version that satisfies the requirement onnxruntime-training==1.17.1 (from versions: 1.15.0, 1.15.1)
ERROR: No matching distribution found for onnxruntime-training==1.17.1

Please install onnxruntime-training using these commands:

python -m pip install cerberus flatbuffers h5py numpy>=1.16.6 onnx packaging protobuf sympy setuptools>=41.4.0
pip install -i https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/ORT/pypi/simple/ onnxruntime-training

We only publish linux wheels for training to PyPI for now.

[elephantpanda]

Hi. Thanks for the reply. I will try it. (it didn't work just installed 1.15.1) I guess also I could create the artifacts in Google Colab too for models that are not too big.

Any reason why only for Linux? I'm guessing this is so it runs on Google Colab?

I am curious about this because I thought OnnxRuntime was a Microsoft project. So a bit strange to publish things to only work on the rival Google platform rather than say Windows.

[GeorgeS2019]

@baijumeswani The link is wrong

pip install -i https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/ORT/pypi/simple/ onnxruntime-training

Example of working link

python -m pip install -i https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-12/pypi/simple/ onnxruntime-gpu