Train PyTorch models with ONNX Runtime

PyTorch/ORT is a Python package that uses ONNX Runtime to accelerate PyTorch model training.

Pre-requisites

You need a machine with at least one NVIDIA or AMD GPU to run PyTorch/ORT.

You can install run PyTorch/ORT in your local environment, or with Docker. If you are using Docker, the following base image is suitable for Nvidia and AMD respectively : nvidia/cuda:11.1.1-cudnn8-devel-ubuntu18.04 or rocm/pytorch:rocm4.1.1_ubuntu18.04_py3.6_pytorch.

Install for Nvidia GPUs

1. Install CUDA

2. Install CuDNN

3. Install PyTorch/ORT and dependencies

Nvidia CUDA version 11.1

• pip install onnx ninja
• pip install --pre torch -f https://download.pytorch.org/whl/nightly/cu111/torch_nightly.html
• pip install --pre onnxruntime-training -f https://onnxruntimepackages.z14.web.core.windows.net/onnxruntime_nightly_cu111.html
• pip install torch-ort

Nvidia CUDA version 10.2

• pip install onnx ninja
• pip install --pre torch -f https://download.pytorch.org/whl/nightly/cu102/torch_nightly.html
• pip install --pre onnxruntime-training -f https://onnxruntimepackages.z14.web.core.windows.net/onnxruntime_nightly_cu102.html
• pip install torch-ort

Install for AMD GPUs

1. Install Rocm 4.1 base package (instructions)

2. Install Rocm 4.1 libraries (instructions)

3. Install Rocm 4.1 RCCL (instructions)

4. Install PyTorch/ORT and dependencies

AMD ROCM version 4.1

• pip install onnx ninja
• pip install --pre torch -f https://download.pytorch.org/whl/nightly/rocm4.1/torch_nightly.html
• pip install --pre onnxruntime-training -f https://onnxruntimepackages.z14.web.core.windows.net/onnxruntime_nightly_rocm41.html
• pip install torch-ort

to install release package of onnxruntime-training:

• pip install onnxruntime-training

Test your installation

1. Clone this repo

• git clone git@github.com:pytorch/ort.git

2. Install extra dependencies

• pip install wget pandas sklearn transformers

3. Run the training script

• python ./ort/tests/bert_for_sequence_classification.py

Add PyTorch/ORT to your PyTorch training script

import onnxruntime
from torch_ort import ORTModule
model = ORTModule(model)
# PyTorch training script follows

Versioning

CUDA

The PyTorch/ORT package was built with CUDA 11.1. If you have a different version of CUDA installed, you should install the CUDA 11.1 toolkit.

This is a limitation that will be removed in the next release.