Muellerzr fix deepspeed (#23657)

* Fix deepspeed recursion

* Better fix

src/transformers/training_args.py

@@ -1708,7 +1708,8 @@ class TrainingArguments:
         """
         requires_backends(self, ["torch"])
         # Make sure `self._n_gpu` is properly setup.
-        _ = self._setup_devices
+        if not hasattr(self, "_n_gpu"):
+            _ = self._setup_devices
         return self._n_gpu
 
     @property