Running torchscript exported model in Triton throws InferenceServerException

[arunsu]

Description
A clear and concise description of what the bug is.
Inferencing a pretrained TorchVision model in Triton 21.02 is throwing following exception. However the same model runs fine on PyTorch 1.7.1

InferenceServerException: PyTorch execute failure: isTensor() INTERNAL ASSERT FAILED at "/opt/tritonserver/include/torch/ATen/core/ivalue_inl.h":152, please report a bug to PyTorch. Expected Tensor but got GenericDict
Exception raised from toTensor at /opt/tritonserver/include/torch/ATen/core/ivalue_inl.h:152 (most recent call first):
frame #0: c10::Error::Error(c10::SourceLocation, std::__cxx11::basic_string<char, std::char_traits, std::allocator >) + 0x6c (0x7fc2e007044c in /opt/tritonserver/backends/pytorch/libc10.so)
frame #1: + 0x8073 (0x7fc2e00a3073 in /opt/tritonserver/backends/pytorch/libtriton_pytorch.so)
frame #2: + 0x15882 (0x7fc2e00b0882 in /opt/tritonserver/backends/pytorch/libtriton_pytorch.so)
frame #3: TRITONBACKEND_ModelInstanceExecute + 0x411 (0x7fc2e00b1d71 in /opt/tritonserver/backends/pytorch/libtriton_pytorch.so)
frame #4: + 0x2e4047 (0x7fc32292b047 in /opt/tritonserver/bin/../lib/libtritonserver.so)
frame #5: + 0xf88a0 (0x7fc32273f8a0 in /opt/tritonserver/bin/../lib/libtritonserver.so)
frame #6: + 0xd6d84 (0x7fc322181d84 in /usr/lib/x86_64-linux-gnu/libstdc++.so.6)
frame #7: + 0x9609 (0x7fc32261c609 in /usr/lib/x86_64-linux-gnu/libpthread.so.0)
frame #8: clone + 0x43 (0x7fc321e6f293 in /usr/lib/x86_64-linux-gnu/libc.so.6)

Triton Information
What version of Triton are you using?
21.02-py3

Are you using the Triton container or did you build it yourself?
nvcr.io/nvidia/tritonserver:21.02-py3

To Reproduce
Steps to reproduce the behavior.
Get the pretrained TorchVision Retinanet model

import torch
import torchvision.models as models
retina50 = models.detection.retinanet_resnet50_fpn(pretrained=True)
retina50.eval()
retina50_scripted = torch.jit.script(retina50)

Test the model with dummy input
dummy_input = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)] # We should run a quick test
scripted_output = retina50_scripted(dummy_input)

Now save this model and run in Triton
retina50_model_scripted.save('retinanet50/1/model.pt')

Describe the models (framework, inputs, outputs), ideally include the model configuration file (if using an ensemble include the model configuration file for that as well).
name: "retinanet50"
platform: "pytorch_libtorch"
input [
{
name: "input__0"
data_type: TYPE_FP32
dims: [3, 480, 640]
}
]
output [
{
name: "output__boxes"
data_type: TYPE_FP32
dims: [93, 4]
},
{
name: "output__scores"
data_type: TYPE_FP32
dims: [93]
},
{
name: "output__labels"
data_type: TYPE_FP32
dims: [93]
}
]

Expected behavior
A clear and concise description of what you expected to happen.
The output from Triton Inference server same as the output of the model run in pytorch.

[CoderHam]

Duplicated by #2593
Closing in favor of #2593