Shuffle based on both the global rank and worker id

src/qusi/internal/light_curve_dataset.py

@@ -40,7 +40,7 @@
 if TYPE_CHECKING:
     from collections.abc import Iterable, Iterator
 
-    from qusi.internal.light_curve_collection import LightCurveObservationCollection
+from qusi.internal.light_curve_collection import LightCurveObservationCollection
 
 
 class OutOfBoundsInjectionHandlingMethod(Enum):
@@ -78,12 +78,27 @@ def __init__(
             raise ValueError(error_message)
         self.post_injection_transform: Callable[[Any], Any] = post_injection_transform
         self.worker_randomizing_set: bool = False
+        self.global_rank: int | None = None
+        self.world_size: int | None = None
+        self.worker_id: int | None = None
+        self.number_of_workers: int | None = None
+        self.global_worker_rank: int | None = None
 
     def __iter__(self):
+        # TODO: This is a hack and there are certainly better places to put this.
         if not self.worker_randomizing_set:
             worker_info = torch.utils.data.get_worker_info()
+            if self.global_rank is None:
+                self.global_rank = 0
+                self.world_size = 1
             if worker_info is not None:
-                self.seed_random(worker_info.id)
+                self.worker_id = worker_info.id
+                self.number_of_workers = worker_info.num_workers
+            else:
+                self.worker_id = 0
+                self.number_of_workers = 0
+            self.global_worker_rank = (self.global_rank * self.number_of_workers) + self.worker_id
+            self.seed_random(self.worker_id)
             self.worker_randomizing_set = True
         base_light_curve_collection_iter_and_type_pairs: list[
             tuple[Iterator[Path], Callable[[Path], LightCurveObservation], LightCurveCollectionType]

src/qusi/internal/lightning_train_session.py

@@ -2,7 +2,6 @@
 
 import datetime
 import logging
-import math
 from pathlib import Path
 from warnings import warn
 
@@ -91,6 +90,10 @@ def train_session(
         accelerator=system_configuration.accelerator,
         logger=loggers,
     )
+    # TODO: Not a fan of needing to magically pass the process number to the datasets here.
+    for train_dataset in train_datasets:
+        train_dataset.global_rank = trainer.global_rank
+        train_dataset.world_size = trainer.world_size
 
     train_dataset = InterleavedDataset.new(*train_datasets)
     workers_per_dataloader = system_configuration.preprocessing_processes_per_train_process