support eager attention

config.json

@@ -30,8 +30,10 @@
   },
   "torch_dtype": "bfloat16",
   "transformers_version": "4.45.0",
+  "_attn_implementation": "flash_attention_2",
   "vision_config": {
     "_flash_attn_2_enabled": true,
+    "_attn_implementation": "flash_attention_2",
     "architectures": [
       "AriaVisionModel"
     ],

modeling_aria.py

@@ -133,6 +133,7 @@ class AriaForConditionalGeneration(AriaPretrainedModel):
     def __init__(self, config: AriaConfig):
         super().__init__(config)
 
+        config.vision_config._attn_implementation = config._attn_implementation
         self.vision_tower = AriaVisionModel(config.vision_config)
         self.multi_modal_projector = build_mm_projector(config)
         self.vocab_size = config.text_config.vocab_size

vision_encoder.py

@@ -38,7 +38,6 @@ class AriaVisionConfig(SiglipVisionConfig):
         **kwargs,
     ):
         super().__init__(**kwargs)
-        self._attn_implementation = "flash_attention_2"
 
 
 class IdentityOp(torch.nn.Module):