rerere hotfix

handler.py

@@ -47,14 +47,14 @@ class EndpointHandler():
         return inputs, True
 
     def _format_inputs(self, inputs: list[str]):
-        prompts = [summary_prompt.format(abstract, "") for abstract in inputs]
+        prompts = [self.summary_prompt.format(abstract, "") for abstract in inputs]
         prompts_lengths = [len(prompt) for prompt in prompts]
         return prompts, prompts_lengths
 
     def _generate_outputs(self, inputs):
-        tokenized = tokenizer(inputs, return_tensors="pt", padding=True).to("cuda")
-        outputs = model.generate(**tokenized, max_new_tokens=500, use_cache=True)
-        decoded = tokenizer.batch_decode(outputs, skip_special_tokens=True) 
+        tokenized = self.tokenizer(inputs, return_tensors="pt", padding=True).to("cuda")
+        outputs = self.model.generate(**tokenized, max_new_tokens=500, use_cache=True)
+        decoded = self.tokenizer.batch_decode(outputs, skip_special_tokens=True) 
         return decoded
 
     def _format_outputs(self, outputs: list[str], inputs_lengths: list[int]):