-
-
Notifications
You must be signed in to change notification settings - Fork 617
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Improve idist to use Tcp and FileStore for init_method #1432
Comments
How about users customize this using environment variables? like this:
in What are the drawbacks of this? |
@ahmedo42 this is interesting idea. Initially, I was thinking about an argument here: ignite/ignite/distributed/comp_models/native.py Lines 78 to 82 in 6d490c8
but maybe we can think about two options like that... |
So the end result would be something like that? def _create_from_backend(self, backend: str, timeout: Optional[int] = None,
init_method:Optional[str] = "env://", **kwargs: Any) -> None:
if backend == dist.Backend.NCCL and not torch.cuda.is_available():
raise RuntimeError("Nccl backend is required but no cuda capable devices")
self.setup_env_vars()
self._local_rank = int(os.environ["LOCAL_RANK"])
# for debug purposes
self._master_port = int(os.environ["MASTER_PORT"]) # type: Optional[int]
self._master_addr = os.environ["MASTER_ADDR"] # type: Optional[str]
init_pg_kwargs = {}
if timeout is not None:
init_pg_kwargs["timeout"] = timeout
if not _valid_init_method(init_method):
#raise error
dist.init_process_group(backend, init_method=init_method, **init_pg_kwargs) |
Yes, something like that. Maybe, |
🚀 Feature
idist
supports default initilializationenv://
. It could be useful to supporttcp://
andfile://
modes (to tackle distributed computations on windows for instance).Main issue is how to configure these modes ?
The text was updated successfully, but these errors were encountered: