I am trying to export a model from PyTorch to Onnx using the new
torch.onnx.dynamo_export interface. Based on this article and the result I got, it seems that graphs exported by
torch.onnx.dynamo_export are composed of aten functions like what is shown in the post here. However, what I want is a graph composed of ONNX ops instead just like what is exported by the old
torch.onnx.export method (i.e., if I keep clicking the
f icon on the top right corner of these aten functions in Netron, I can get the decomposed graph with only ONNX ops I want). Just wondering, is there a way to achieve a flattened ONNX graph without these hierarchy of aten funcs to ONNX ops with the new dynamo export interface?