prim::ListConstruct fusion error

[ulyssesxxxi]

Hi,

I'm trying to compile maskrcnn-benchmark with pytorch/tvm (directly from the master branch), and even though I know there are currently several operators that are yet to be implemented, I wanted to advance as much as possible before they are made available, so I'm reporting this in case the issue is unrelated.

I'm starting from the following PR from @t-vi that implements several missing pieces needed for JIT in maskrcnn-benchmark: facebookresearch/maskrcnn-benchmark. But after adding torch_tvm.enable() to trace_model.py, it throws an error during trace checking:

RuntimeError: input->node()->kind() == prim::Constant INTERNAL ASSERT FAILED at ../torch/csrc/jit/passes/graph_fuser.cpp:359, please report a bug to PyTorch. (mergeNodeIntoGroup at ../torch/csrc/jit/passes/graph_fuser.cpp:359)

I've found that the offending Node kind() is prim::ListConstruct, but I don't know how to proceed any further. My bet is some issue with BoxList (the return type of the model), but I'm not sure how to confirm this point. Any suggestion is sincerely welcome.

Thx in advance.