[QUERY] How to stop torch-mlir from decomposing certain ops when using torch dynamo

[Abhishek-TyRnT]

When I was torchscript.compile, there is a argument backend_legal_ops which optionally stops torch-mlir from decomposing certain aten ops. But I couldn't find something similar when using fx.import_and_export. If there is someone please help how to do it.

[mgehre-amd]

import_and_export has a decomposition_table argument. By default, it uses the table given by get_decomposition_table. You can try explicitly passing an empty list to disable all decompositions at this level.
(Other decompositions might still happen due to how dynamo is working internally).

[Abhishek-TyRnT]

I am trying to stop decomposition of operations like layer_norm or softmax. Even after passing the empty list, it decomposing. Although I found an alternate route to do that. like torch script, we can use backend_legal_ops argument before running the pass like this.

torch-mlir/projects/pt1/python/torch_mlir/torchscript.py

Lines 339 to 346 in 11efcda

	option_string = (
	"{backend-legal-ops="
	+ ",".join(backend_legal_ops)
	+ " extra-library="
	+ extra_library_file_name
	+ "}"
	)
	run_pipeline_with_repro_report(

This seems like a useful feature, and since torch-mlir is migrating more towards dynamo, I believe dynamo export must also have this feature. I can send a PR for this