Apply style and revert change in sft_video_llm example (#4214)

Author: qgallouedec

examples/scripts/reward_modeling.py

@@ -67,7 +67,6 @@
     get_kbit_device_map,
     get_peft_config,
     get_quantization_config,
-    setup_chat_format,
 )
 
 

examples/scripts/sft_gpt_oss.py

@@ -75,7 +75,6 @@ def main(script_args, training_args, model_args):
 
     model = AutoModelForCausalLM.from_pretrained(model_args.model_name_or_path, **model_kwargs)
 
-
     # Load dataset
     dataset = load_dataset(script_args.dataset_name, name=script_args.dataset_config)
 

examples/scripts/sft_video_llm.py

@@ -62,7 +62,7 @@
 from datasets import load_dataset
 from peft import LoraConfig
 from qwen_vl_utils import process_vision_info
-from transformers import AutoModelForImageTextToText, BitsAndBytesConfig, Qwen2VLProcessor
+from transformers import AutoModelForImageTextToText, AutoProcessor, BitsAndBytesConfig, Qwen2VLProcessor
 
 from trl import ModelConfig, ScriptArguments, SFTConfig, SFTTrainer, TrlParser, get_kbit_device_map
 
@@ -224,6 +224,10 @@ class CustomScriptArguments(ScriptArguments):
         model.config.use_reentrant = False
         model.enable_input_require_grads()
 
+    processor = AutoProcessor.from_pretrained(
+        model_args.model_name_or_path, trust_remote_code=model_args.trust_remote_code
+    )
+
     # Prepare dataset
     prepared_dataset = [prepare_dataset(example, script_args.video_cache_dir) for example in dataset]
 
@@ -234,6 +238,7 @@ class CustomScriptArguments(ScriptArguments):
         train_dataset=prepared_dataset,
         data_collator=collate_fn,
         peft_config=peft_config,
+        processing_class=processor,
     )
 
     # Train model

trl/scripts/dpo.py

@@ -81,7 +81,6 @@
     get_peft_config,
     get_quantization_config,
 )
-from trl.trainer.utils import SIMPLE_CHAT_TEMPLATE
 
 
 logger = logging.get_logger(__name__)