Initialize single-process distributed group for single GPU inference compatibility

sample.py

@@ -54,7 +54,29 @@ DEFAULT_CAPTIONS = {
 
 def setup_model_and_components(args: argparse.Namespace) -> Tuple[torch.nn.Module, Optional[torch.nn.Module], tuple]:
     """Initialize and load the model, VAE, and text encoder."""
+    # Initialize distributed training context
+    # For single GPU inference, we still need to initialize process group
+    # because the model code uses torch.distributed.all_reduce
     dist = utils.Distributed()
+    
+    # If not running with torchrun, initialize single-process group
+    if not dist.distributed and torch.cuda.is_available():
+        import os
+        # Initialize single-process process group for model compatibility
+        if not torch.distributed.is_initialized():
+            os.environ['MASTER_ADDR'] = 'localhost'
+            os.environ['MASTER_PORT'] = '12355'
+            os.environ['RANK'] = '0'
+            os.environ['LOCAL_RANK'] = '0'
+            os.environ['WORLD_SIZE'] = '1'
+            torch.distributed.init_process_group(
+                backend='nccl' if torch.cuda.is_available() else 'gloo',
+                init_method='env://',
+                world_size=1,
+                rank=0,
+            )
+            print("✅ Initialized single-process distributed group for model compatibility")
+    
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
     # Set random seed