Fix set of model parallel in the Trainer when no GPUs are available (#25239)

src/transformers/trainer.py

@@ -378,10 +378,13 @@ class Trainer:
             devices = [device for device in set(model.hf_device_map.values()) if device not in ["cpu", "disk"]]
             if len(devices) > 1:
                 self.is_model_parallel = True
-            else:
+            elif len(devices) == 1:
                 self.is_model_parallel = self.args.device != torch.device(devices[0])
+            else:
+                self.is_model_parallel = False
 
             # warn users
+            if self.is_model_parallel:
                 logger.info(
                     "You have loaded a model on multiple GPUs. `is_model_parallel` attribute will be force-set"
                     " to `True` to avoid any unexpected behavior such as device placement mismatching."