Upgrade perf_run script to support TRT 10 and fix some issues

[zewenli98]

Description

1. removed embedding layer from constant folding lowering pass
2. upgraded TRT API to execute_async_v3
3. added optimization_level arg and set highest optimization 5 as default
4. fixed some known issues

Fixes #3634

Type of change

• Bug fix (non-breaking change which fixes an issue)

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/py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py	2025-07-17 21:15:44.135432+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py	2025-07-17 21:16:11.238491+00:00
@@ -3596,6 +3596,6 @@
        SourceIR.ATEN,
        name,
        input=args[0],
        weight=args[1],
        bias=args_bounds_check(args, 2, None),
-    )
\ No newline at end of file
+    )
--- /home/runner/work/TensorRT/TensorRT/tools/perf/perf_run.py	2025-07-17 21:15:44.172433+00:00
+++ /home/runner/work/TensorRT/TensorRT/tools/perf/perf_run.py	2025-07-17 21:16:16.387016+00:00
@@ -776,11 +776,13 @@
        raise ValueError(
            "No valid models specified. Please provide a torchscript model file or model name (defined in hub.py) or model_hf name in huggingface models "
        )

    backends = parse_backends(params["backends"])
-    if any(backend in ["dynamo", "torch_compile", "tensorrt"] for backend in backends) and (model_torch is None):
+    if any(
+        backend in ["dynamo", "torch_compile", "tensorrt"] for backend in backends
+    ) and (model_torch is None):
        raise ValueError(
            "No Pytorch model (nn.Module) is provided for torchdynamo compilation. Please provide a pytorch model using --model_torch argument"
        )

    batch_size = params["batch_size"]

[peri044]

1. remove lower_linear
2. consider try/catch for onnx export
3. Update perf_run readme with the latest arguments
4. Run llm example with the constant_folding to test accuracy

[peri044]

modify the start timer to include creation of input bindings as well in run_tensorrt

[HolyWu]

I think that registering a converter for OpOverloadPacket has no effect.

[zewenli98]

I found for some models in fp16, for example, bert, registering a linear op can reduce latency. It seems no effect for fp32 though.

[HolyWu]

torch_tensorrt.dynamo.types