torch-ort 1.17.0

Accelerate PyTorch models with ONNX Runtime

The torch-ort packages uses the PyTorch APIs to accelerate PyTorch models using ONNX Runtime.

Dependencies

The torch-ort package depends on the onnxruntime-training package, which depends on specific versions of GPU libraries such as NVIDIA CUDA.

The default command pip install torch-ort installs the onnxruntime-training version that depends on CUDA 10.2.

If you have a different version of CUDA installed, you can install a different version of onnxruntime-training explicitly:

• CUDA 11.1 pip install onnxruntime-training -f https://download.onnxruntime.ai/onnxruntime_stable_cu111.html

Post-installation step

Once torch-ort is installed, there is a post-installation step:

python -m torch_ort.configure

If this step fails, it is likely due to GPU library version mismatch between onnxruntime-training and your installation. You can check the version of onnxruntime-training by running pip list. For example:

onnxruntime-training 1.9.0+cu111

Releases

• 1.9.0

  Release Notes : https://github.com/pytorch/ort/releases/tag/v1.9.0

• 1.8.1

  Release Notes : https://github.com/pytorch/ort/releases/tag/v1.8.1