[brief] Add a way to set the sharing strategy from torch.multiprocess…

…ing.

[detailed]

src/helios/trainer.py

@@ -225,6 +225,7 @@ def __init__(
         src_root: pathlib.Path | None = None,
         import_prefix: str = "",
         print_banner: bool = True,
+        mp_sharing_strategy: str | None = None,
     ):
         """Create the trainer."""
         self._model: hlm.Model | None = None
@@ -347,6 +348,21 @@ def test(self, model: hlm.Model, datamodule: data.DataModule) -> None:
                 logging.close_default_loggers()
             raise RuntimeError("error: uncaught exception") from e
 
+    def set_mp_sharing_strategy(self, strategy: str) -> None:
+        """
+        Set the torch multiprocessing strategy.
+
+        This can be used to switch the sharing strategy used by torch.multiprocessing, but
+        it must be called prior to any calls to fit or test.
+        Note:
+            If the trainer is set to not use distributed training or if it's started from
+            torchrun, then this function does nothing.
+        """
+        if not self._is_distributed or self._is_torchrun:
+            return
+
+        mp.set_sharing_strategy(strategy)
+
     def _launch(
         self, model: hlm.Model, datamodule: data.DataModule, mode: _TrainerMode
     ) -> None: