feat: custom binding names for out of runtime deployment

[narendasan]

Description

Allows users of the convert_exported_program_to_serialized_engine API to specify binding names for all bindings utilizing the exported programs pytrees for i/o flattening to flatten the provided names and associate names to specific io positions.

TODO: Add support for aliasing inputs and outputs through this method (see: #4251)

Fixes #4212

Type of change

Please delete options that are not relevant and/or add your own.

• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

Checklist:

• My code follows the style guidelines of this project (You can use the linters)
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas and hacks
• I have made corresponding changes to the documentation
• I have added tests to verify my fix or my feature
• New and existing unit tests pass locally with my changes
• I have added the relevant labels to my PR in so that relevant reviewers are notified

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/examples/dynamo/engine_converter_binding_names.py	2026-05-29 19:38:55.472711+00:00
+++ /home/runner/work/TensorRT/TensorRT/examples/dynamo/engine_converter_binding_names.py	2026-05-29 19:39:15.963114+00:00
@@ -302,11 +302,11 @@

try:
    torch_tensorrt.dynamo.convert_exported_program_to_serialized_trt_engine(
        exported,
        arg_inputs=(x,),
-        output_binding_names=("only_one",),   # wrong arity for the 2-tuple return
+        output_binding_names=("only_one",),  # wrong arity for the 2-tuple return
        require_full_compilation=True,
        min_block_size=1,
        use_python_runtime=False,
        immutable_weights=True,
    )
--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/_TRTInterpreter.py	2026-05-29 19:38:55.486463+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/_TRTInterpreter.py	2026-05-29 19:39:17.826414+00:00
@@ -397,11 +397,10 @@
        run_module_start_time = datetime.now()
        super().run()
        _LOGGER.info(
            f"TRT INetwork construction elapsed time: {datetime.now() - run_module_start_time}"
        )
-

    @staticmethod
    def find_weight(
        weight_name: str,
        weight_refit_map: dict[str, Any],
--- /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_engine_converter_binding_names.py	2026-05-29 19:38:55.514192+00:00
+++ /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_engine_converter_binding_names.py	2026-05-29 19:39:21.059130+00:00
@@ -12,10 +12,11 @@
We ``tree_flatten`` each and verify spec equality against the exported
program's specs.  On success the flat lists concatenate (args then
kwargs) into FX's flattened placeholder order; the interpreter just
indexes positionally.
"""
+
import unittest

import tensorrt as trt
import torch
import torch_tensorrt
@@ -361,10 +362,12 @@
        (input_name,) = self._binding_names(engine, trt.TensorIOMode.INPUT)
        outs = self._run_engine(engine, {input_name: x.contiguous()})
        with torch.no_grad():
            ref = model(x)
        torch.testing.assert_close(outs["p_out"], ref["primary"], rtol=1e-2, atol=1e-2)
-        torch.testing.assert_close(outs["s_out"], ref["secondary"], rtol=1e-2, atol=1e-2)
+        torch.testing.assert_close(
+            outs["s_out"], ref["secondary"], rtol=1e-2, atol=1e-2
+        )


if __name__ == "__main__":
    run_tests()

[lanluo-nvidia]

LGTM