Add support for pytorch 1.5.0

[Giuseppe5]

As per #119
Initial commit, more to follow.

[volcacius]

Currently conda dry runs for retrieving cache hashes are failing, because the session is not defined in Nox for 1.5.0. You need to add "1.5.0" to PYTORCH_VERSIONS in gen_github_actions.py, that's the single source of truth that Nox adheres too.

[volcacius]

Both Windows and Ubuntu are failing in ways I currently don't understand. Can you try disabling extension building with ninja? Add a use_ninja=False flag to the options here

brevitas/setup.py

Line 180 in 24d0275

'build_ext': BuildJittableExtension.with_options(no_python_abi_suffix=True),

[Giuseppe5]

That is one solution, another one proposed in the changelog is the following:
If a previously succeeding python setup.py install now fails, try setting the MAX_JOBS environment variable.

I will try both and see which one is better (faster, easier to configure, and everything).

[Giuseppe5]

I'll leave it here for memo. I need to refactor the torch.autograd functions that are used when the JIT is disabled, since there was a change in the API. This means also a change in the setup.py file for the codegen. Hopefully the new APIs will work also with older version (we were using deprecated APIs, even though no warning was ever printed about that).
If that's not the case, then we'll need support for both version according to pytorch version.

Edit: I was mistaken, we are using the correct version.

[Giuseppe5]

Now everything works as expected, and the errors on MacOS are unrelated to brevitas-support for 1.5

[volcacius]

Move these two constants into brevitas/__init__.py where they are also used, so that we have a single source of truth. init.py is not importing anything new compared to setup.py so it should be fine

[volcacius]

Also, for coherence it should be MAX_TORCH_JITTABLE_VERSION = 1.4.0, since the MIN one is inclusive, and then adjust comparison operators in init and setup accordingly

[Giuseppe5]

In init.py there's the import of packaging that breaks everything. Considering the workaround we have in setup.py for having it installed automatically, I would prefer duplicating the two constants for now and look for a cleaner solution in a second moment

[volcacius]

Just move the import inside the run() methods and then pass the value to the template function.

[volcacius]

Actually we can just define a bool is_ste_jittable in __init__.py and then import it into setup.py from within the run methods.

[Giuseppe5]

Importing within the run methods causes all the code in init to be executed, included the search for the "_C" file. But at setup-time, the _C module has yet to be created, so it fails.

[volcacius]

That's a good point, let's put the bool in brevitas.config and import it into both init and setup. We need to setup a separate pull to move docstrings out of config first though, and into brevitas' init, or we will have another setup time import error.

[volcacius]

Also please rebase into two commits, one "Setup: add support for Pytorch 1.5.0" with setup.py and init.py, the other "CI/CD: run CI against Pytorch 1.5.0" with everything else

[volcacius]

Please rebase after #136 .

[Giuseppe5]

No fix seems to be working for MacOS, pytorch1.5 and python 3.6.
This (open) issue seems related, even though it happens on ubuntu (in our case, only with MacOS)