Fix: num_workers=0 to avoid multiprocessing crash with few shards

scripts/training/train_flux_lora.py

@@ -210,9 +210,8 @@ def main():
     # Dataset
     print(f"  Loading dataset from {args.data_dir}")
     train_dataset, num_shards = create_webdataset(args.data_dir, args.resolution, args.batch_size)
-    num_workers = min(2, num_shards)
     train_dataloader = torch.utils.data.DataLoader(
-        train_dataset, batch_size=None, num_workers=num_workers, pin_memory=True
+        train_dataset, batch_size=None, num_workers=0
     )
 
     # Training loop