ONNX ops for training with PyTorch and TensorFlow

[kiszk]

I am excited to see activities to enable model training on ONNX. How is ONNX going to co-operate with other AI frameworks (e.g. PyTorch, TensorFlow) in a context of training? The current ORT seems to convert an inference graph from PyTorch to a training one by adding operations, such as loss functions, backward computations, and optimizers, as described in this presentation.
Is this approach along with a long-term plan for ORT? Or is ORT expected to consume a training graph from PyTorch or TensorFlow, for forward/backward/update, using operations in com.microsoft domain ops for training?

Please let me know if there is a better forum to discuss this topic.
Thank you so much.