resolving zero3 init when using accelerate config with Trainer (#25227)

* resolving zero3 init when using accelerate config with Trainer

* refactor

* fix

* fix import

src/transformers/training_args.py

@@ -1627,6 +1627,12 @@ class TrainingArguments:
 
             os.environ["ACCELERATE_USE_DEEPSPEED"] = "true"
             self.deepspeed_plugin = DeepSpeedPlugin(hf_ds_config=self.hf_deepspeed_config)
+        elif strtobool(os.environ.get("ACCELERATE_USE_DEEPSPEED", "false")):
+            # Accelerate DeepSpeed Plugin
+            from accelerate.utils import DeepSpeedPlugin
+
+            self.deepspeed_plugin = DeepSpeedPlugin()
+            self.deepspeed_plugin.set_deepspeed_weakref()
 
         if self.push_to_hub_token is not None:
             warnings.warn(