[RAG] Fix rag from pretrained question encoder generator behavior (#11962)

* fix_torch_device_generate_test

* remove @

* fix rag from pretrained loading

* add test

* uplaod

* finish

tests/test_modeling_rag.py

@@ -1132,12 +1132,17 @@ class RagModelSaveLoadTests(unittest.TestCase):
                 "facebook/bart-large-cnn",
                 retriever=rag_retriever,
                 config=rag_config,
+                question_encoder_max_length=200,
+                generator_max_length=200,
             ).to(torch_device)
             # check that the from pretrained methods work
             rag_token.save_pretrained(tmp_dirname)
             rag_token.from_pretrained(tmp_dirname, retriever=rag_retriever)
             rag_token.to(torch_device)
 
+            self.assertTrue(rag_token.question_encoder.config.max_length == 200)
+            self.assertTrue(rag_token.generator.config.max_length == 200)
+
             with torch.no_grad():
                 output = rag_token(
                     input_ids,