Fix FA2 tests (#29909)

* fix FA2 tests

* refactor inference test name

tests/models/bark/test_modeling_bark.py

@@ -879,7 +879,7 @@ def test_resize_embeddings_untied(self):
     @require_torch_gpu
     @pytest.mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference(self):
+    def test_flash_attn_2_inference_equivalence(self):
         for model_class in self.all_model_classes:
             if not model_class._supports_flash_attn_2:
                 return
@@ -936,7 +936,7 @@ def test_flash_attn_2_inference(self):
     @require_torch_gpu
     @pytest.mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference_padding_right(self):
+    def test_flash_attn_2_inference_equivalence_right_padding(self):
         for model_class in self.all_model_classes:
             if not model_class._supports_flash_attn_2:
                 return

tests/models/distilbert/test_modeling_distilbert.py

@@ -301,7 +301,7 @@ def test_torchscript_device_change(self):
     @require_torch_accelerator
     @pytest.mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference(self):
+    def test_flash_attn_2_inference_equivalence(self):
         import torch
 
         for model_class in self.all_model_classes:
@@ -353,7 +353,7 @@ def test_flash_attn_2_inference(self):
     @require_torch_accelerator
     @pytest.mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference_padding_right(self):
+    def test_flash_attn_2_inference_equivalence_right_padding(self):
         import torch
 
         for model_class in self.all_model_classes:

tests/models/gemma/test_modeling_gemma.py

@@ -462,7 +462,7 @@ def test_flash_attn_2_generate_use_cache(self):
     @require_torch_gpu
     @pytest.mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference_padding_right(self):
+    def test_flash_attn_2_inference_equivalence_right_padding(self):
         self.skipTest("Gemma flash attention does not support right padding")
 
     @require_torch_sdpa

tests/models/mistral/test_modeling_mistral.py

@@ -466,7 +466,7 @@ def test_flash_attn_2_generate_use_cache(self):
     @require_torch_gpu
     @pytest.mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference_padding_right(self):
+    def test_flash_attn_2_inference_equivalence_right_padding(self):
         self.skipTest("Mistral flash attention does not support right padding")
 
 

tests/models/mixtral/test_modeling_mixtral.py

@@ -465,7 +465,7 @@ def test_flash_attn_2_generate_use_cache(self):
     @require_torch_gpu
     @pytest.mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference_padding_right(self):
+    def test_flash_attn_2_inference_equivalence_right_padding(self):
         self.skipTest("Mixtral flash attention does not support right padding")
 
     # Ignore copy

tests/models/qwen2/test_modeling_qwen2.py

@@ -477,7 +477,7 @@ def test_flash_attn_2_generate_use_cache(self):
     @require_torch_gpu
     @pytest.mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference_padding_right(self):
+    def test_flash_attn_2_inference_equivalence_right_padding(self):
         self.skipTest("Qwen2 flash attention does not support right padding")
 
 

tests/models/starcoder2/test_modeling_starcoder2.py

@@ -461,7 +461,7 @@ def test_flash_attn_2_generate_use_cache(self):
     @require_torch_gpu
     @pytest.mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference_padding_right(self):
+    def test_flash_attn_2_inference_equivalence_right_padding(self):
         self.skipTest("Starcoder2 flash attention does not support right padding")
 
 

tests/models/whisper/test_modeling_whisper.py

@@ -888,7 +888,7 @@ def _check_outputs(self, output, input_ids, config, use_cache=False, num_return_
     @require_torch_gpu
     @pytest.mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference(self):
+    def test_flash_attn_2_inference_equivalence(self):
         import torch
 
         for model_class in self.all_model_classes:
@@ -934,7 +934,7 @@ def test_flash_attn_2_inference(self):
     @require_torch_gpu
     @pytest.mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference_padding_right(self):
+    def test_flash_attn_2_inference_equivalence_right_padding(self):
         import torch
 
         for model_class in self.all_model_classes:

tests/test_modeling_common.py

@@ -3245,7 +3245,7 @@ def test_flash_attn_2_conversion(self):
     @require_torch_gpu
     @mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference(self):
+    def test_flash_attn_2_inference_equivalence(self):
         for model_class in self.all_model_classes:
             if not model_class._supports_flash_attn_2:
                 self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
@@ -3260,9 +3260,7 @@ def test_flash_attn_2_inference(self):
                 )
                 model_fa.to(torch_device)
 
-                model = model_class.from_pretrained(
-                    tmpdirname, torch_dtype=torch.bfloat16, attn_implementation="flash_attention_2"
-                )
+                model = model_class.from_pretrained(tmpdirname, torch_dtype=torch.bfloat16)
                 model.to(torch_device)
 
                 dummy_input = inputs_dict[model.main_input_name][:1]
@@ -3340,7 +3338,7 @@ def test_flash_attn_2_inference(self):
     @require_torch_gpu
     @mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference_padding_right(self):
+    def test_flash_attn_2_inference_equivalence_right_padding(self):
         for model_class in self.all_model_classes:
             if not model_class._supports_flash_attn_2:
                 self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
@@ -3355,9 +3353,7 @@ def test_flash_attn_2_inference_padding_right(self):
                 )
                 model_fa.to(torch_device)
 
-                model = model_class.from_pretrained(
-                    tmpdirname, torch_dtype=torch.bfloat16, attn_implementation="flash_attention_2"
-                )
+                model = model_class.from_pretrained(tmpdirname, torch_dtype=torch.bfloat16)
                 model.to(torch_device)
 
                 dummy_input = inputs_dict[model.main_input_name][:1]