🚀 Feature / 🐛 Bug

Ops which are bound in at torch/functional.py and torch/tensor.py are only exposed to the python api and not JIT / C++. Example issues #20113 and #29340.

An example PR fix is here: #28918. Another fix may be fixing resolution for JIT so that it tries to compile these functions.

Edit: the stack in #33737 fixed the functions in torch/functional.py with an exception or two

Ops bound in torch/functional.py:

- lu_unpack
- isinf
- stft
- unique
- unique_consecutive
- tensordot
- cdist
- norm
- lu

Ops bound in torch/tensor.py:

• norm -
• register_hook - not currently possible, first class functions
• retain_grad
• is_shared
• share_memory_
• __reversed__
• lu
• stft
• split
• unique
• unique_consecutive
• refine_names
• align_to
• unflatten
• rename_
• update_names_
• __contains__

Another set of ops not supported is in torch.nn.init. The full list can be seen here: https://pytorch.org/docs/master/jit_unsupported.html

cc @gmagogsfm @suo