Update handler.py

handler.py

@@ -15,10 +15,11 @@ class EndpointHandler():
         self.device = "cuda" if torch.cuda.is_available() else "cpu"
         self.model_base = "Salesforce/blip2-opt-2.7b"
         self.model_name = "sooh-j/blip2-vizwizqa"
-        self.base_model = Blip2ForConditionalGeneration.from_pretrained(self.model_base, load_in_8bit=True)
+        # self.base_model = Blip2ForConditionalGeneration.from_pretrained(self.model_base, load_in_8bit=True)
         # self.pipe = Blip2ForConditionalGeneration.from_pretrained(self.model_base, load_in_8bit=True, torch_dtype=torch.float16)
-        self.processor = Blip2Processor.from_pretrained(self.base_model_name)
-        self.model = PeftModel.from_pretrained(self.model_name, self.base_model_name).to(self.device)
+        self.processor = Blip2Processor.from_pretrained(self.model_name)
+        self.model = Blip2ForConditionalGeneration.from_pretrained(self.model_name).to(self.device)
+        # self.model = PeftModel.from_pretrained(self.model_name, self.base_model_name).to(self.device)
 
         self.device = "cuda" if torch.cuda.is_available() else "cpu"
         self.model.to(self.device)