Update handler.py

handler.py

@@ -31,7 +31,7 @@ class EndpointHandler:
         self.model = AutoModelForVision2Seq.from_pretrained(
             model_dir,
             torch_dtype=torch.bfloat16,
-            _attn_implementation="flash_attention_2",
+            _attn_implementation="eager", # "flash_attention_2",
             device_map="auto",
         ).eval()
         self.generation_config = GenerationConfig.from_pretrained(model_dir)