Rename _supports_flash_attn_2 in examples and tests (#39471)

* delete `_supports_flash_attn_2` from examples and tests

* simplify docs

examples/modular-transformers/modeling_multimodal2.py

@@ -497,8 +497,7 @@ class Multimodal2VisionPreTrainedModel(PreTrainedModel):
     base_model_prefix = "multimodal2_vision"
     supports_gradient_checkpointing = True
     _supports_sdpa = True
-    _supports_flash_attn_2 = True
-    _supports_flash_attn_3 = True
+    _supports_flash_attn = True
     _supports_flex_attn = True
     _supports_attention_backend = True
 

examples/modular-transformers/modeling_my_new_model2.py

@@ -289,8 +289,7 @@ class MyNewModel2PreTrainedModel(PreTrainedModel):
     supports_gradient_checkpointing = True
     _no_split_modules = ["MyNewModel2DecoderLayer"]
     _skip_keys_device_placement = ["past_key_values"]
-    _supports_flash_attn_2 = True
-    _supports_flash_attn_3 = True
+    _supports_flash_attn = True
     _supports_sdpa = True
     _supports_flex_attn = True
     _supports_cache_class = True

examples/modular-transformers/modeling_new_task_model.py

@@ -95,8 +95,7 @@ class NewTaskModelPreTrainedModel(PreTrainedModel):
     _supports_cache_class = True
     _supports_quantized_cache = True
     _supports_static_cache = True
-    _supports_flash_attn_2 = True
-    _supports_flash_attn_3 = True
+    _supports_flash_attn = True
     _supports_sdpa = True
     _supports_flex_attn = True
     _supports_attention_backend = True

examples/modular-transformers/modeling_super.py

@@ -288,8 +288,7 @@ class SuperPreTrainedModel(PreTrainedModel):
     supports_gradient_checkpointing = True
     _no_split_modules = ["SuperDecoderLayer"]
     _skip_keys_device_placement = ["past_key_values"]
-    _supports_flash_attn_2 = True
-    _supports_flash_attn_3 = True
+    _supports_flash_attn = True
     _supports_sdpa = True
     _supports_flex_attn = True
     _supports_cache_class = True