Update handler.py

handler.py

@@ -25,10 +25,10 @@ class EndpointHandler():
         inputs = data.pop("inputs", data)
         parameters = data.pop("parameters", {})
 
-        parameters["max_new_tokens"] = parameters.pop("max_new_tokens", DEFAULT_MAX_NEW_TOKENS)
+        parameters["max_tokens"] = parameters.pop("max_tokens", DEFAULT_MAX_NEW_TOKENS)
         logger.info(inputs)
         logger.info(parameters)
-        if parameters["max_new_tokens"] > MAX_MAX_NEW_TOKENS:
+        if parameters["max_tokens"] > MAX_MAX_NEW_TOKENS:
             logger.error(f"requested max_new_tokens too high (> {MAX_MAX_NEW_TOKENS})")
             return [{"generated_text": None, "error": f"requested max_new_tokens too high (> {MAX_MAX_NEW_TOKENS})"}]