exclude fsdp from delay_optimizer_creation

src/transformers/trainer.py

@@ -2172,7 +2172,7 @@ def _inner_training_loop(
             else:
                 debug_overflow = DebugUnderflowOverflow(self.model)  # noqa
 
-        delay_optimizer_creation = is_sagemaker_mp_enabled() or self.is_fsdp_xla_enabled or self.is_fsdp_enabled
+        delay_optimizer_creation = is_sagemaker_mp_enabled() or self.is_fsdp_xla_enabled
 
         # We need to reset the scheduler, as its parameters may be different on subsequent calls
         if self._created_lr_scheduler:
@@ -2221,9 +2221,12 @@ def _inner_training_loop(
         # FSDP-XLA, SageMaker MP/DP, DataParallel, IPEX
         use_accelerator_prepare = True if model is self.model else False
 
+        # configure fsdp plugin for qlora if any
+        if use_accelerator_prepare:
+            self._fsdp_qlora_plugin_updates()
+
         if delay_optimizer_creation:
             if use_accelerator_prepare:
-                self._fsdp_qlora_plugin_updates()
                 self.model = self.accelerator.prepare(self.model)
             self.create_optimizer_and_scheduler(num_training_steps=max_steps)