[distributed] multiprocessing.spawn python version check

[teng-li]

This will be super helpful to the user

[pietern]

The version check already happens in __init__.py. You can move this assertion into the else branch of that version check and define a spawn function there that just raises.

[pietern]

The else branch of this one:

pytorch/torch/multiprocessing/__init__.py

Lines 37 to 40 in b8de8f6

	if sys.version_info >= (3, 4):
	"""Add helper function to spawn N processes and wait for completion of any of
	them. This depends `mp.get_context` which was added in Python 3.4."""
	from .spawn import spawn, SpawnContext

[teng-li]

@pietern Not a big fan of creating another dummy class for this purpose. Changed to check the version at runtime and we always import these classes.

[pietern]

I didn't imply another class, just:

def spawn(*args, **kwargs):
  raise RuntimeError("Python >= 3.4 required bla bla bla")

[facebook-github-bot]

@teng-li is landing this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[pietern]

Here's the conundrum: nobody should be using the SpawnContext directly, but its join function should show up in the docs, because that's how you end up doing checking, waiting, and finally ensuring that all processes terminate. Mentioning the existence of some unusable SpawnContext in the error message only adds confusion IMO.

Two possible solutions:

1. No more SpawnContext: This would be some anonymous object with a join function, that we can somehow include in the documentation (no idea how).
2. More SpawnContext: Make it usable by itself and turn spawn into a simple factory function.

The latter is a bit better because we don't have to worry about the docs and the class is usable by itself.

Filed #14047. No need to block this.