Gate deep imports from torch.distributed (#13673)

Author: hlky

src/diffusers/training_utils.py

@@ -14,10 +14,15 @@
 import torch.nn.functional as F
 
 
-if getattr(torch, "distributed", None) is not None:
+if torch.distributed.is_available():
     from torch.distributed.fsdp import CPUOffload, ShardingStrategy
     from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
     from torch.distributed.fsdp.wrap import transformer_auto_wrap_policy
+else:
+    CPUOffload = None
+    ShardingStrategy = None
+    FSDP = None
+    transformer_auto_wrap_policy = None
 
 from .models import UNet2DConditionModel
 from .pipelines import DiffusionPipeline