fix #6391

.gitignore

@@ -172,6 +172,3 @@ saves/
 output/
 wandb/
 generated_predictions.jsonl
-
-# unittest
-dummy_dir/

src/llamafactory/hparams/generating_args.py

@@ -15,6 +15,8 @@
 from dataclasses import asdict, dataclass, field
 from typing import Any, Dict, Optional
 
+from transformers import GenerationConfig
+
 
 @dataclass
 class GeneratingArguments:
@@ -69,10 +71,17 @@ class GeneratingArguments:
         metadata={"help": "Whether or not to remove special tokens in the decoding."},
     )
 
-    def to_dict(self) -> Dict[str, Any]:
+    def to_dict(self, obey_generation_config: bool = False) -> Dict[str, Any]:
         args = asdict(self)
         if args.get("max_new_tokens", -1) > 0:
             args.pop("max_length", None)
         else:
             args.pop("max_new_tokens", None)
+
+        if obey_generation_config:
+            generation_config = GenerationConfig()
+            for key in list(args.keys()):
+                if not hasattr(generation_config, key):
+                    args.pop(key)
+
         return args

src/llamafactory/train/sft/trainer.py

@@ -151,7 +151,7 @@ def _pad_tensors_to_target_len(self, src_tensor: "torch.Tensor", tgt_tensor: "to
         return padded_tensor.contiguous()  # in contiguous memory
 
     def save_predictions(
-        self, dataset: "Dataset", predict_results: "PredictionOutput", gen_kwargs: Dict[str, Any]
+        self, dataset: "Dataset", predict_results: "PredictionOutput", skip_special_tokens: bool = True
     ) -> None:
         r"""
         Saves model predictions to `output_dir`.
@@ -179,12 +179,8 @@ def save_predictions(
                 preds[i] = np.concatenate((preds[i][pad_len[0] :], preds[i][: pad_len[0]]), axis=-1)
 
         decoded_inputs = self.processing_class.batch_decode(dataset["input_ids"], skip_special_tokens=False)
-        decoded_preds = self.processing_class.batch_decode(
-            preds, skip_special_tokens=gen_kwargs["skip_special_tokens"]
-        )
-        decoded_labels = self.processing_class.batch_decode(
-            labels, skip_special_tokens=gen_kwargs["skip_special_tokens"]
-        )
+        decoded_preds = self.processing_class.batch_decode(preds, skip_special_tokens=skip_special_tokens)
+        decoded_labels = self.processing_class.batch_decode(labels, skip_special_tokens=skip_special_tokens)
 
         with open(output_prediction_file, "w", encoding="utf-8") as f:
             for text, pred, label in zip(decoded_inputs, decoded_preds, decoded_labels):

src/llamafactory/train/sft/workflow.py

@@ -91,7 +91,7 @@ def run_sft(
     )
 
     # Keyword arguments for `model.generate`
-    gen_kwargs = generating_args.to_dict()
+    gen_kwargs = generating_args.to_dict(obey_generation_config=True)
     gen_kwargs["eos_token_id"] = [tokenizer.eos_token_id] + tokenizer.additional_special_tokens_ids
     gen_kwargs["pad_token_id"] = tokenizer.pad_token_id
     gen_kwargs["logits_processor"] = get_logits_processor()
@@ -130,7 +130,7 @@ def run_sft(
             predict_results.metrics.pop("predict_loss", None)
         trainer.log_metrics("predict", predict_results.metrics)
         trainer.save_metrics("predict", predict_results.metrics)
-        trainer.save_predictions(dataset_module["eval_dataset"], predict_results, gen_kwargs)
+        trainer.save_predictions(dataset_module["eval_dataset"], predict_results, generating_args.skip_special_tokens)
 
     # Create model card
     create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)

tests/e2e/test_train.py

@@ -60,12 +60,12 @@
     ],
 )
 def test_run_exp(stage: str, dataset: str):
-    output_dir = os.path.join("output", f"dummy_dir/train_{stage}")
+    output_dir = os.path.join("output", f"train_{stage}")
     run_exp({"stage": stage, "dataset": dataset, "output_dir": output_dir, **TRAIN_ARGS})
     assert os.path.exists(output_dir)
 
 
 def test_export():
-    export_dir = os.path.join("output", "dummy_dir/llama3_export")
+    export_dir = os.path.join("output", "llama3_export")
     export_model({"export_dir": export_dir, **INFER_ARGS})
     assert os.path.exists(export_dir)

tests/train/test_sft_trainer.py

@@ -58,7 +58,11 @@ def __call__(self, features: List[Dict[str, Any]]) -> Dict[str, Any]:
 @pytest.mark.parametrize("disable_shuffling", [False, True])
 def test_shuffle(disable_shuffling: bool):
     model_args, data_args, training_args, finetuning_args, _ = get_train_args(
-        {"output_dir": f"dummy_dir/{disable_shuffling}", "disable_shuffling": disable_shuffling, **TRAIN_ARGS}
+        {
+            "output_dir": os.path.join("output", f"shuffle{str(disable_shuffling).lower()}"),
+            "disable_shuffling": disable_shuffling,
+            **TRAIN_ARGS,
+        }
     )
     tokenizer_module = load_tokenizer(model_args)
     tokenizer = tokenizer_module["tokenizer"]