[shoumikhin]

What is wrong today

torch_tensorrt.save() re-exports a compiled module by default (retrace=True), and re-exporting needs example inputs. So the shortest way to write an ExecuTorch file fails:

trt_gm = torch_tensorrt.compile(model, ir="dynamo", inputs=inputs)
torch_tensorrt.save(trt_gm, "model.pte", output_format="executorch")
# ValueError: ... retrace is True, however no inputs specs were provided ...

To make it work, users add retrace=False, and often arg_inputs, to every call. The ExecuTorch guide and its examples all do this. torch_tensorrt.executorch.export() already treats an unset retrace as False, so the two entry points disagree.

What this change does

• retrace now defaults to None. save() picks the value from the output format: False for "executorch", True for every other format, as before.
• An explicit retrace=True or retrace=False still wins.
• The ExecuTorch save() snippets in the docs drop retrace=False and arg_inputs.

After this change:

torch_tensorrt.save(trt_gm, "model.pte", output_format="executorch")

Callers that pass inputs and relied on the old default now take the retrace=False path. For a module from torch_tensorrt.compile() this writes the same .pte, and the existing warning says the inputs are not needed. A GraphModule that did not go through Torch-TensorRT compilation still needs retrace=True.

Tests

• Two new unit tests in tests/py/dynamo/executorch/test_api.py. The first saves a GraphModule to ExecuTorch with neither retrace nor inputs and checks it takes the no-retrace path. It fails without this change. The second checks that other formats still retrace by default.
• tests/py/dynamo/executorch and test_exporter_inlining.py: same results before and after, apart from the new test.
• End to end on one H100 with nightly wheels: static and dynamic shapes, with and without inputs, and with an explicit retrace=True. Each .pte loads in the ExecuTorch runtime and matches eager output. With inputs, the old and new code write the same file.
• The four ExecuTorch examples (static shape, dynamic shape, TensorRT plus CUDA, device resident) run with no retrace and no inputs.

[cehongwang]

LGTM