Somehow in the new torchvision version, the behavior of loading googlenet is different when no weights are specified (leads to segfaults when compiling via migraphx, not sure why but that can be a different issue to look at). Works as expected once pretrained weights are specified.

The tracer change is because the GoogLeNet model is loaded with empty modules (see below snippet):

   (branch4): Sequential(
      (0): MaxPool2d(kernel_size=3, stride=1, padding=1, dilation=1, ceil_mode=True)
      (1): BasicConv2d(
        (conv): Conv2d(832, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn): BatchNorm2d(128, eps=0.001, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
  )
  (aux1): None
  (aux2): None
  (avgpool): AdaptiveAvgPool2d(output_size=(1, 1))

In this case those modules need to be skipped by the tracer (not sure why they exist all of a sudden in the new version).