Make training args fully immutable (#25435)

* Make training args fully immutable

* Working tests, PyTorch

* In test_trainer

* during testing

* Use proper dataclass way

* Fix test

* Another one

* Fix tf

* Lingering slow

* Exception

* Clean

examples/pytorch/summarization/run_summarization.py

@@ -18,6 +18,7 @@ Fine-tuning the library models for sequence to sequence.
 """
 # You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.
 
+import dataclasses
 import logging
 import os
 import sys
@@ -674,14 +675,10 @@ def main():
         return result
 
     # Override the decoding parameters of Seq2SeqTrainer
-    training_args.generation_max_length = (
-        training_args.generation_max_length
-        if training_args.generation_max_length is not None
-        else data_args.val_max_target_length
-    )
-    training_args.generation_num_beams = (
-        data_args.num_beams if data_args.num_beams is not None else training_args.generation_num_beams
-    )
+    if training_args.generation_max_length is None:
+        training_args = dataclasses.replace(training_args, generation_max_length=data_args.val_max_target_length)
+    if training_args.generation_num_beams is None:
+        training_args = dataclasses.replace(training_args, generation_num_beams=data_args.num_beams)
 
     # Initialize our Trainer
     trainer = Seq2SeqTrainer(