How do I get all the required inputs when using export_params=False in torch.onnx.export + onnx-mlir

[srcarroll]

I'm using torch.onnx.export(..., export_params=False) to avoid inlining all the weights of the model as constants in the IR. For the most part this works as expected, i.e. i get all the model parameters as inputs in the main func.func signature. However, for some models there are extra.

I've figured out that some of the extras are from model.buffers(). I can even get parameters or buffers by name since the func has func arg attributes like onnx.name = "L__self___transformer_h_0_attn_c_attn_bias". But then there are the outliers I can't figure out how to get. For example I get an extra arg that has attribute onnx.name = "onnx::MatMul_1270". Where do args like this come from? Is there a simple way to extract all needed inputs (in the right order) from either the model object or the fx graph?

Any guidance is appreciated. Thanks!