Fix flash attention bugs with Mistral and Falcon (#27625)

* fix various bugs with flash attention

* bump

* fix test

* fix mistral

* use skiptest instead of return that may be misleading

* fix on review

src/transformers/models/falcon/modeling_falcon.py

@@ -564,6 +564,12 @@ class FalconFlashAttention2(FalconAttention):
 
         past_key_value = (key_layer, value_layer) if use_cache else None
 
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_layer = query_layer.transpose(1, 2)
+        key_layer = key_layer.transpose(1, 2)
+        value_layer = value_layer.transpose(1, 2)
+
         if alibi is not None:
             raise ValueError("`alibi` is not supported when `use_flash_attn` is True")
 

src/transformers/models/mistral/modeling_mistral.py

@@ -838,7 +838,7 @@ class MistralModel(MistralPreTrainedModel):
             attention_mask is not None
             and hasattr(self.config, "_flash_attn_2_enabled")
             and self.config._flash_attn_2_enabled
-            and past_key_values is not None
+            and use_cache
         ):
             is_padding_right = attention_mask[:, -1].sum().item() != batch_size
             if is_padding_right:

tests/models/llama/test_modeling_llama.py

@@ -22,6 +22,7 @@ from parameterized import parameterized
 
 from transformers import LlamaConfig, is_torch_available, set_seed
 from transformers.testing_utils import (
+    require_bitsandbytes,
     require_flash_attn,
     require_torch,
     require_torch_accelerator,
@@ -385,6 +386,7 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
 
     @require_flash_attn
     @require_torch_gpu
+    @require_bitsandbytes
     @pytest.mark.flash_attn_test
     @slow
     def test_flash_attn_2_generate_padding_right(self):

tests/models/mistral/test_modeling_mistral.py

@@ -375,9 +375,6 @@ class MistralModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMi
         import torch
 
         for model_class in self.all_generative_model_classes:
-            if not model_class._supports_flash_attn_2:
-                return
-
             config, _ = self.model_tester.prepare_config_and_inputs_for_common()
             model = model_class(config)
 
@@ -405,36 +402,49 @@ class MistralModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMi
     @require_torch_gpu
     @pytest.mark.flash_attn_test
     @slow
-    def test_flash_attn_2_inference_padding_right(self):
+    def test_flash_attn_2_generate_use_cache(self):
         import torch
 
-        for model_class in self.all_model_classes:
+        max_new_tokens = 30
-            if not model_class._supports_flash_attn_2:
+
-                return
+        for model_class in self.all_generative_model_classes:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+            dummy_input = inputs_dict[model_class.main_input_name]
+            if dummy_input.dtype in [torch.float32, torch.bfloat16]:
+                dummy_input = dummy_input.to(torch.float16)
+
+            # make sure that all models have enough positions for generation
+            if hasattr(config, "max_position_embeddings"):
+                config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
 
-            config, _ = self.model_tester.prepare_config_and_inputs_for_common()
             model = model_class(config)
 
             with tempfile.TemporaryDirectory() as tmpdirname:
                 model.save_pretrained(tmpdirname)
-                model_fa = model_class.from_pretrained(
+
-                    tmpdirname, torch_dtype=torch.bfloat16, use_flash_attention_2=True
+                dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
-                )
+                # NOTE: Mistral apparently does not support right padding + use_cache with FA2.
-                model_fa.to(torch_device)
+                dummy_attention_mask[:, -1] = 1
 
                 model = model_class.from_pretrained(
-                    tmpdirname, torch_dtype=torch.bfloat16, use_flash_attention_2=False
+                    tmpdirname,
+                    torch_dtype=torch.float16,
+                    use_flash_attention_2=True,
+                    low_cpu_mem_usage=True,
+                ).to(torch_device)
+
+                # Just test that a large cache works as expected
+                _ = model.generate(
+                    dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
                 )
-                model.to(torch_device)
 
-                dummy_input = torch.LongTensor([[1, 2, 3, 4, 5]]).to(torch_device)
+    @require_flash_attn
-                dummy_attention_mask = torch.LongTensor([[1, 1, 1, 1, 0]]).to(torch_device)
+    @require_torch_gpu
-
+    @pytest.mark.flash_attn_test
-                _ = model(dummy_input, output_hidden_states=True).hidden_states[-1]
+    @slow
-                with self.assertRaises(ValueError):
+    def test_flash_attn_2_inference_padding_right(self):
-                    _ = model_fa(
+        self.skipTest("Mistral flash attention does not support right padding")
-                        dummy_input, attention_mask=dummy_attention_mask, output_hidden_states=True
-                    ).hidden_states[-1]
 
 
 @require_torch

tests/test_modeling_common.py

@@ -2835,7 +2835,7 @@ class ModelTesterMixin:
 
         for model_class in self.all_model_classes:
             if not model_class._supports_flash_attn_2:
-                return
+                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
 
             model = model_class(config)
 
@@ -2860,7 +2860,7 @@ class ModelTesterMixin:
 
         for model_class in self.all_model_classes:
             if not model_class._supports_flash_attn_2:
-                return
+                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
 
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
             model = model_class(config)
@@ -2957,7 +2957,7 @@ class ModelTesterMixin:
 
         for model_class in self.all_model_classes:
             if not model_class._supports_flash_attn_2:
-                return
+                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
 
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
             model = model_class(config)
@@ -3050,7 +3050,7 @@ class ModelTesterMixin:
 
         for model_class in self.all_generative_model_classes:
             if not model_class._supports_flash_attn_2:
-                return
+                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
 
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
             model = model_class(config)
@@ -3093,7 +3093,7 @@ class ModelTesterMixin:
 
         for model_class in self.all_generative_model_classes:
             if not model_class._supports_flash_attn_2:
-                return
+                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
 
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
             model = model_class(config)
@@ -3109,7 +3109,7 @@ class ModelTesterMixin:
                     dummy_input = dummy_input.to(torch.float16)
 
                 dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
-                # make sure we do left padding
+                # make sure we do right padding
                 dummy_attention_mask[:, :-1] = 1
                 dummy_attention_mask[:, -1:] = 0
 
@@ -3138,7 +3138,7 @@ class ModelTesterMixin:
 
         for model_class in self.all_generative_model_classes:
             if not model_class._supports_flash_attn_2:
-                return
+                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
 
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 
@@ -3179,7 +3179,7 @@ class ModelTesterMixin:
 
         for model_class in self.all_generative_model_classes:
             if not model_class._supports_flash_attn_2:
-                return
+                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
 
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
             model = model_class(config)
@@ -3279,7 +3279,7 @@ class ModelTesterMixin:
 
         for model_class in self.all_generative_model_classes:
             if not model_class._supports_flash_attn_2:
-                return
+                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
 
             config, _ = self.model_tester.prepare_config_and_inputs_for_common()
             # TODO: to change it in the future with other relevant auto classes