diff --git a/src/transformers/testing_utils.py b/src/transformers/testing_utils.py
index 7bb2d5049dccf8..2781e9e102e050 100644
--- a/src/transformers/testing_utils.py
+++ b/src/transformers/testing_utils.py
@@ -144,6 +144,7 @@
 
 if is_accelerate_available():
     from accelerate.state import AcceleratorState, PartialState
+    from accelerate.utils.imports import is_fp8_available
 
 
 if is_pytest_available():
@@ -1000,6 +1001,13 @@ def require_torch_fp16(test_case):
     )(test_case)
 
 
+def require_fp8(test_case):
+    """Decorator marking a test that requires supports for fp8"""
+    return unittest.skipUnless(is_accelerate_available() and is_fp8_available(), "test requires fp8 support")(
+        test_case
+    )
+
+
 def require_torch_bf16(test_case):
     """Decorator marking a test that requires a device that supports bf16"""
     return unittest.skipUnless(
diff --git a/src/transformers/trainer.py b/src/transformers/trainer.py
index 1b13787007e9c3..7f9f5ec76f42e2 100755
--- a/src/transformers/trainer.py
+++ b/src/transformers/trainer.py
@@ -2209,7 +2209,7 @@ 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:
@@ -2258,9 +2258,12 @@ 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)
 
diff --git a/tests/trainer/test_trainer_fsdp.py b/tests/trainer/test_trainer_fsdp.py
index 994a82a8db0c44..4bcf5de04520e2 100644
--- a/tests/trainer/test_trainer_fsdp.py
+++ b/tests/trainer/test_trainer_fsdp.py
@@ -20,6 +20,8 @@
     execute_subprocess_async,
     get_torch_dist_unique_port,
     require_accelerate,
+    require_fp8,
+    require_fsdp,
     require_torch_multi_gpu,
 )
 
@@ -64,6 +66,7 @@ def __getitem__(self, i: int) -> str:
 class TestFSDPTrainer(TestCasePlus):
     @require_accelerate
     @require_torch_multi_gpu
+    @require_fsdp
     def test_trainer(self):
         output_dir = self.get_auto_remove_tmp_dir()
         cmd = [
@@ -86,6 +89,35 @@ def test_trainer(self):
         # successful return here == success - any errors would have caused an error in the sub-call
 
 
+class TestFSDPTrainerFP8(TestCasePlus):
+    @require_accelerate
+    @require_torch_multi_gpu
+    @require_fsdp
+    @require_fp8
+    def test_trainer(self):
+        output_dir = self.get_auto_remove_tmp_dir()
+        cmd = [
+            "accelerate",
+            "launch",
+            "--use_fsdp",
+            "--main_process_port",
+            f"{get_torch_dist_unique_port()}",
+            "--num_processes",
+            f"{torch.cuda.device_count()}",
+            "--mixed_precision",
+            "fp8",
+            "--fsdp_transformer_layer_cls_to_wrap",
+            "GPT2Block",
+            f"{self.test_file_dir}/test_trainer_fsdp.py",
+            "--output_dir",
+            f"{output_dir}",
+            "--report_to",
+            "none",
+        ]
+        execute_subprocess_async(cmd, env=self.get_env())
+        # successful return here == success - any errors would have caused an error in the sub-call
+
+
 if __name__ == "__main__":
     parser = HfArgumentParser((Seq2SeqTrainingArguments,))
     training_args = parser.parse_args_into_dataclasses()[0]