fix(test): enable TRT-RTX refit and engine cache tests

Now that #4181 removed the RTX-specific batch norm workaround that
bypassed constant folding, the refit bug (#3752) is resolved — eps
constants are no longer created as separate CONSTANT layers on RTX.

Remove the RTX skip decorators from:
- test_dynamo_compile_with_refittable_weight_stripped_engine
- test_dynamo_compile_with_custom_engine_cache
- test_dynamo_compile_change_input_shape

Keep the RTX skip on test_caching_small_model, which fails a timing
assertion (cached compilation is slower than uncached on RTX). Update
the skip message to reflect the actual reason.

Fix import ordering in test_weight_stripped_engine.py: tensorrt must
be imported after torch_tensorrt so the tensorrt_rtx module alias is
resolved correctly.

Fixes #3752

Author: tp5uiuc

tests/py/dynamo/models/test_engine_cache.py

@@ -268,11 +268,6 @@ def remove_timing_cache(path=TIMING_CACHE_PATH):
     @unittest.skipIf(
         not importlib.util.find_spec("torchvision"), "torchvision not installed"
     )
-    @unittest.skipIf(
-        torch_trt.ENABLED_FEATURES.tensorrt_rtx,
-        # TODO: need to fix this https://github.com/pytorch/TensorRT/issues/3752
-        "There is bug in refit, so we skip the test for now",
-    )
     def test_dynamo_compile_with_custom_engine_cache(self):
         model = models.resnet18(pretrained=True).eval().to("cuda")
 
@@ -342,11 +337,6 @@ def test_dynamo_compile_with_custom_engine_cache(self):
     @unittest.skipIf(
         not importlib.util.find_spec("torchvision"), "torchvision not installed"
     )
-    @unittest.skipIf(
-        torch_trt.ENABLED_FEATURES.tensorrt_rtx,
-        # TODO: need to fix this https://github.com/pytorch/TensorRT/issues/3752
-        "There is bug in refit, so we skip the test for now",
-    )
     def test_dynamo_compile_change_input_shape(self):
         """Runs compilation 3 times, the cache should miss each time"""
         model = models.resnet18(pretrained=True).eval().to("cuda")
@@ -659,8 +649,7 @@ def forward(self, c, d):
     )
     @unittest.skipIf(
         torch_trt.ENABLED_FEATURES.tensorrt_rtx,
-        # TODO: need to fix this https://github.com/pytorch/TensorRT/issues/3752
-        "There is bug in refit, so we skip the test for now",
+        "Engine caching compilation time assertion is unreliable with TensorRT-RTX",
     )
     def test_caching_small_model(self):
         from torch_tensorrt.dynamo._refit import refit_module_weights

tests/py/dynamo/models/test_weight_stripped_engine.py

@@ -4,7 +4,6 @@
 import shutil
 import unittest
 
-import tensorrt as trt
 import torch
 import torch_tensorrt as torch_trt
 from torch.testing._internal.common_utils import TestCase
@@ -13,6 +12,8 @@
 from torch_tensorrt.dynamo._refit import refit_module_weights
 from torch_tensorrt.dynamo.utils import COSINE_THRESHOLD, cosine_similarity
 
+import tensorrt as trt  # isort: skip  # must import after torch_tensorrt to resolve tensorrt_rtx alias
+
 assertions = unittest.TestCase()
 
 if importlib.util.find_spec("torchvision"):
@@ -272,11 +273,6 @@ def test_engine_caching_saves_weight_stripped_engine(self):
         not importlib.util.find_spec("torchvision"),
         "torchvision is not installed",
     )
-    @unittest.skipIf(
-        torch_trt.ENABLED_FEATURES.tensorrt_rtx,
-        # TODO: need to fix this https://github.com/pytorch/TensorRT/issues/3752
-        "There is bug in refit, so we skip the test for now",
-    )
     def test_dynamo_compile_with_refittable_weight_stripped_engine(self):
         pyt_model = models.resnet18(pretrained=True).eval().to("cuda")
         example_inputs = (torch.randn((100, 3, 224, 224)).to("cuda"),)