fix ci

Author: yuanlehome

llm/predict/predictor.py

@@ -1200,15 +1200,17 @@ def create_predictor(
 
     max_position_embeddings = get_model_max_position_embeddings(config)
     if max_position_embeddings is None:
-        max_position_embeddings = 2048
-        logger.warning("Can not retrieval `max_position_embeddings` from config.json, use default value 2048")
-
-    if predictor_args.src_length + predictor_args.max_length > max_position_embeddings:
-        raise ValueError(
-            f"The sum of src_length<{predictor_args.src_length}> and "
-            f"max_length<{predictor_args.max_length}> should be smaller than or equal to "
-            f"the maximum position embedding size<{max_position_embeddings}>"
+        max_position_embeddings = predictor_args.src_length + predictor_args.max_length
+        logger.warning(
+            f"Can not retrieval `max_position_embeddings` from config.json, use default value {max_position_embeddings}"
         )
+    else:
+        if predictor_args.src_length + predictor_args.max_length > max_position_embeddings:
+            raise ValueError(
+                f"The sum of src_length<{predictor_args.src_length}> and "
+                f"max_length<{predictor_args.max_length}> should be smaller than or equal to "
+                f"the maximum position embedding size<{max_position_embeddings}>"
+            )
 
     # update config parameter for inference predictor
     if predictor_args.decode_strategy == "greedy_search":

tests/llm/test_predictor.py

@@ -17,7 +17,6 @@
 import unittest
 
 import paddle
-import pytest
 from parameterized import parameterized_class
 
 from paddlenlp.experimental.transformers import QWenForQWenVLInferenceModel
@@ -199,28 +198,6 @@ def test_predictor(self):
         self.assertGreaterEqual(count / len(result_0), 0.8)
 
 
-class PredictorBaseTest(LLMTest, unittest.TestCase):
-    def load_test_config(self):
-        config = load_test_config("./tests/fixtures/llm/predictor.yaml", "inference-predict")
-        config["model_name_or_path"] = "__internal_testing__/micro-random-llama"
-
-        return config
-
-    def test_create_predictor_with_unexpected_length(self):
-        from predict.predictor import predict
-
-        config = self.load_test_config()
-        config.pop("src_length", None)
-        config.pop("max_length", None)
-
-        with pytest.raises(ValueError, match="The sum of src_length<1025> and"):
-            config["max_length"] = 1024
-            config["src_length"] = 1025
-
-            with argv_context_guard(config):
-                predict()
-
-
 @parameterized_class(
     ["model_name_or_path", "model_class"],
     [