ddp fixes for training (#22874)

ddp fixes for stable lm training

src/transformers/trainer.py

@@ -1565,12 +1565,13 @@ def patched_optimizer_step(optimizer, barrier=False, optimizer_args={}):
                 kwargs["bucket_cap_mb"] = self.args.ddp_bucket_cap_mb
             if is_torch_neuroncore_available():
                 return model
-            model = nn.parallel.DistributedDataParallel(
-                model,
-                device_ids=[self.args.local_rank] if self.args._n_gpu != 0 else None,
-                output_device=self.args.local_rank if self.args._n_gpu != 0 else None,
-                **kwargs,
-            )
+            if any(p.requires_grad for p in model.parameters()):
+                model = nn.parallel.DistributedDataParallel(
+                    model,
+                    device_ids=[self.args.local_rank] if self.args._n_gpu != 0 else None,
+                    output_device=self.args.local_rank if self.args._n_gpu != 0 else None,
+                    **kwargs,
+                )
 
         # torch.compile() needs to be called after wrapping the model with FSDP or DDP
         # to ensure that it accounts for the graph breaks required by those wrappers
@@ -1920,6 +1921,7 @@ def _inner_training_loop(
                     (total_batched_samples % args.gradient_accumulation_steps != 0)
                     and args.parallel_mode == ParallelMode.DISTRIBUTED
                     and args._no_sync_in_gradient_accumulation
+                    and hasattr(model, "no_sync")
                 ):
                     # Avoid unnecessary DDP synchronization since there will be no backward pass on this example.
                     with model.no_sync():