With this PR, we removed an redundant API torch.export.export_for_inference().
If you are using it and affected by this BC-breaking change, please replace your callsite in the following way
Before:
ep = export_for_inference(mod, args, kwargs, dynamic_shapes=dynamic_shape_spec, strict=strict_flag, preserve_module_call_signature=signature_spec, decomp_table=decomp_rules)
After:
ep = export_for_training(mod, args, kwargs, dynamic_shapes=dynamic_shape_spec, strict=strict_flag, preserve_module_call_signature=signature_spec)
ep = ep.run_decompositions(decomp_table=decomp_rules)