↔ [Converter] Add Dynamo converters for all Elementwise ATen Operators

[gs-olive]

Context

The existing FX operators use different implementations for the elementwise operators than Dynamo does. Specifically, many networks mix Int and Float types for additions, subtractions, multiplications, etc. These tend to cause errors when using the FX converters.

Proposal

Reimplement a light wrapper over each of the elementwise operators in ATen to use the Dynamo implementation of ATen, featured in #2070. Update relevant imports.

Once implemented, remove the following:

TensorRT/tests/py/dynamo/converters/test_binary_ops_aten.py

Line 80 in 7d9705d

@unittest.skip("Pending reimplementation of all binary converters in Dynamo")

[narendasan]

@gs-olive What in py/torch_tensorrt/dynamo/conversion/impl/elementwise needs to be changed to address this?

[gs-olive]

For each of the available elementwise ops, there should be an implementation/decorator in py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py. The body of each function should generally only require calls to the convert_binary_elementwise utility in the Dynamo directory, with the appropriate operation type.

If a binary operator requires a more complex operation, py/torch_tensorrt/dynamo/conversion/impl/elementwise/ops.py should be modified first.

[github-actions]

This issue has not seen activity for 90 days, Remove stale label or comment or this will be closed in 10 days

[gs-olive]

Completed on main