🐛 [Bug] Encountered bug when using TRTModuleNext in Dynamo

[gs-olive]

Bug Description

When compiling the fasterrcnn_mobilenet_v3_large_320_fpn with torch_tensorrt.dynamo.compile, using the use_experimental_rt=True argument, the following error is encountered:

  Outputs: [
    id: 0
      name: output0
      shape: [2, 960, 13, 14]
      dtype: Float
    id: 1
      name: output1
      shape: [1, 160, 1, 1]
      dtype: Float
    id: 2
      name: output2
      shape: [1, 960, 1, 1]
      dtype: Float
    id: 3
      name: output3
      shape: [2, 960, 1, 1]
      dtype: Float
  }

...

DEBUG: [Torch-TensorRT - Debug Build] - Output Name: output0 Shape: [2, 960, 13, 14]
DEBUG: [Torch-TensorRT - Debug Build] - Output Name: output0 Shape: [2, 960, 13, 14]
DEBUG: [Torch-TensorRT - Debug Build] - Output Name: output0 Shape: [2, 960, 13, 14]
DEBUG: [Torch-TensorRT - Debug Build] - Output Name: output3 Shape: [2, 960, 1, 1]
ERROR: [Torch-TensorRT - Debug Build] - 3: [executionContext.cpp::enqueueV3::2666] Error Code 3: API Usage Error (Parameter check failed at: runtime/api/executionContext.cpp::enqueueV3::2666, condition: mContext.profileObliviousBindings.at(profileObliviousIndex) || getPtrOrNull(mOutputAllocators, profileObliviousIndex)

Note the difference in the logged outputs and the engine-reported outputs.

To Reproduce

Steps to reproduce the behavior:

model = torchvision.models.detection.fasterrcnn_mobilenet_v3_large_320_fpn().eval().cuda()
inp = torch.rand((3, 300, 400)).cuda()
inp2 = torch.rand((3, 500, 400)).cuda()
model_acc = torch_tensorrt.dynamo.compile(model, [inp, inp2], pass_through_build_failures=True, use_experimental_rt=True)

Expected behavior

Model should compile without errors.

Environment

• Torch-TensorRT Version (e.g. 1.0.0): 075a028
• PyTorch Version (e.g. 1.0): 2.1.0.dev20230606+cu118

Additional context

Related: #1565, #1995