🐛 [Bug] Fallback models with multiple outputs and encounter bug during inference

[cyfwry]

Bug Description

Compilation can be done when I fallback a model with multiple outputs. However, the following problems will occur when doing inference with the optimized model:

Traceback (most recent call last):
  File "test_insert.py", line 35, in <module>
    r(torch.ones([1,12,224,224]).cuda())
  File "/usr/local/lib64/python3.6/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
RuntimeError: Method (but not graphs in general) require a single output. Use None/Tuple for 0 or 2+ outputs

To Reproduce

Just run:

import torch
import torch.nn as nn
from pylib_torch import torch_tensorrt as tt

class Model(nn.Module):
    def __init__(self):
        super(Model,self).__init__()
        self.conv1=nn.Conv2d(3,3,3)
        self.conv2=nn.Conv2d(3,3,3)
        self.conv3=nn.Conv2d(3,3,3)
        self.conv4=nn.Conv2d(3,3,3)

    def forward(self,x):
        x1=x[:,0:3]
        x2=x[:,3:6]
        x3=x[:,6:9]
        x4=x[:,9:12]
        out1=self.conv1(x1)
        out2=self.conv2(x2)
        out3=self.conv3(x3)
        out4=self.conv4(x4)
        return out1,out2,out3,out4

a=Model().cuda().eval()
b=torch.jit.trace(a,torch.ones([1,12,20,20]).cuda())

torch.jit.save(b,'model.ts')
compile_settings = {}
compile_settings["inputs"] = [tt.Input(shape = [1,12,20,20])]
compile_settings["torch_executed_ops"]=['aten::slice']

r=tt.compile(b,**compile_settings)
r(torch.ones([1,12,224,224]).cuda())

Expected behavior

The model compiles and infers correctly.

Environment

Build information about Torch-TensorRT can be found by turning on debug messages

• Torch-TensorRT Version : 1.0.0
• PyTorch Version: 1.10.0
• CPU Architecture: Intel(R) Xeon(R) Platinum 8352Y CPU @ 2.20GHz
• OS (e.g., Linux): CentOS 7
• How you installed PyTorch (conda, pip, libtorch, source): pip
• Build command you used (if compiling from source):
• Are you using local sources or building from archives:
• Python version: 3.6.8
• CUDA version: 11.4
• GPU models and configuration: A30
• Any other relevant information:

Additional context

I have located this bug in torch_tensorrt::core::ConstructFallbackGraph and fixed it locally. After confirming the existence of this bug, I will open a PR to fix it.