Update handler.py

handler.py

@@ -45,7 +45,10 @@ class EndpointHandler():
 
         # Forward
         LOGGER.info(f"Start generation.")
-        generated_ids = self.model.generate(**model_inputs, max_new_tokens=9999999, do_sample=True, pad_token_id=self.tokenizer.eos_token_id)
+        eos_tok = self.tokenizer.eos_token_id
+        LOGGER.info(f"Generating Ids")
+        generated_ids = self.model.generate(**model_inputs, max_new_tokens=9999999, do_sample=True, pad_token_id=eos_tok)
+        LOGGER.info(f"Ids Generated.")
         decoded = self.tokenizer.batch_decode(generated_ids)
         LOGGER.info(f"Generated text length: {len(decoded[0])}")
         return {"generated_text": decoded[0]}