diff --git a/tests/slow/test_grpo_slow.py b/tests/slow/test_grpo_slow.py
index 047c3f9c04d..2e53abb4393 100644
--- a/tests/slow/test_grpo_slow.py
+++ b/tests/slow/test_grpo_slow.py
@@ -208,14 +208,14 @@ def test_training_with_transformers_paged(self, model_name):
 
         release_memory(model, trainer)
 
-    @require_flash_attn
-    @require_bitsandbytes
-    @require_peft
     @parameterized.expand(
         [
             ("HuggingFaceTB/SmolVLM-Instruct",),  # Only test the smaller model to avoid OOM
         ]
     )
+    @require_flash_attn
+    @require_bitsandbytes
+    @require_peft
     def test_vlm_training(self, model_name):
         """
         Test VLM training with aggressive memory optimization.
diff --git a/tests/test_dpo_trainer.py b/tests/test_dpo_trainer.py
index 7e0401a1d0c..fd4f7efb4d7 100644
--- a/tests/test_dpo_trainer.py
+++ b/tests/test_dpo_trainer.py
@@ -1303,7 +1303,6 @@ def test_train_with_length_desensitization(self):
             if param.sum() != 0:  # ignore 0 biases
                 self.assertFalse(torch.allclose(param, new_param, rtol=1e-12, atol=1e-12))
 
-    @unittest.skipUnless(sys.version_info >= (3, 10), "Liger kernel is not supported on Python 3.9")
     @parameterized.expand(
         [
             (0.1, "sigmoid"),
@@ -1319,6 +1318,7 @@ def test_train_with_length_desensitization(self):
         ]
     )
     @require_liger_kernel
+    @unittest.skipUnless(sys.version_info >= (3, 10), "Liger kernel is not supported on Python 3.9")
     def test_dpo_trainer_with_liger(self, beta, loss_type):
         """Test DPO trainer with Liger loss enabled across supported loss types.
 
diff --git a/tests/test_grpo_trainer.py b/tests/test_grpo_trainer.py
index a839a654bca..03d44e50603 100644
--- a/tests/test_grpo_trainer.py
+++ b/tests/test_grpo_trainer.py
@@ -1513,14 +1513,14 @@ def reward_func(completions, **kwargs):
             new_param = trainer.model.get_parameter(n)
             self.assertFalse(torch.equal(param, new_param), f"Parameter {n} has not changed.")
 
-    @require_vision
-    @require_vllm
     @parameterized.expand(
         [
             ("trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration",),
             ("trl-internal-testing/tiny-Gemma3ForConditionalGeneration",),
         ]
     )
+    @require_vision
+    @require_vllm
     @unittest.skip("We should add a mock for the vLLM server.")
     def test_training_vlm_and_vllm(self, model_id) -> None:
         dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
diff --git a/tests/test_online_dpo_trainer.py b/tests/test_online_dpo_trainer.py
index 336db8a089f..b9dec1135e6 100644
--- a/tests/test_online_dpo_trainer.py
+++ b/tests/test_online_dpo_trainer.py
@@ -238,8 +238,8 @@ def test_training_with_peft_model_and_peft_config(self):
         # Check if training loss is available
         self.assertIn("train_loss", trainer.state.log_history[-1])
 
-    @require_llm_blender
     @parameterized.expand([("standard_prompt_only",), ("conversational_prompt_only",)])
+    @require_llm_blender
     def test_training_with_judge(self, config_name):
         training_args = OnlineDPOConfig(
             output_dir=self.tmp_dir,
@@ -419,8 +419,8 @@ def test_generation_config_setup(self):
         self.assertEqual(trainer.generation_config.max_new_tokens, 64)
         self.assertFalse(trainer.generation_config.do_sample)  # From generation_kwargs
 
-    @require_torch_accelerator
     @parameterized.expand([("standard_prompt_only",), ("conversational_prompt_only",)])
+    @require_torch_accelerator
     def test_training_with_transformers_paged(self, config_name):
         if Version(transformers.__version__) < Version("4.57.0"):
             pytest.xfail("Upstream bug in transformers (GH#40692). Fix merged; awaiting release >= 4.57.0")
diff --git a/tests/test_rloo_trainer.py b/tests/test_rloo_trainer.py
index cde52de6047..a2b1d3bf8b7 100644
--- a/tests/test_rloo_trainer.py
+++ b/tests/test_rloo_trainer.py
@@ -1254,14 +1254,14 @@ def reward_func(completions, **kwargs):
             new_param = trainer.model.get_parameter(n)
             self.assertFalse(torch.equal(param, new_param), f"Parameter {n} has not changed.")
 
-    @require_vision
-    @require_vllm
     @parameterized.expand(
         [
             ("trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration",),
             ("trl-internal-testing/tiny-Gemma3ForConditionalGeneration",),
         ]
     )
+    @require_vision
+    @require_vllm
     @unittest.skip("We should add a mock for the vLLM server.")
     def test_training_vlm_and_vllm(self, model_id) -> None:
         dataset = load_dataset("trl-internal-testing/zen-image", "conversational_prompt_only", split="train")
diff --git a/tests/test_xpo_trainer.py b/tests/test_xpo_trainer.py
index 9d50b542a03..9af803830cf 100644
--- a/tests/test_xpo_trainer.py
+++ b/tests/test_xpo_trainer.py
@@ -184,8 +184,8 @@ def test_training_pre_pefted_model_implicit_ref(self):
 
         self.assertIn("train_loss", trainer.state.log_history[-1])
 
-    @require_llm_blender
     @parameterized.expand([("standard_prompt_only",), ("conversational_prompt_only",)])
+    @require_llm_blender
     def test_xpo_trainer_judge_training(self, config_name):
         training_args = XPOConfig(
             output_dir=self.tmp_dir,