Hotfix: Fall back to config.text_config._name_or_path if missing config._name_or_path (#4324)

Author: albertvillanova

trl/trainer/base_trainer.py

@@ -17,7 +17,7 @@
 
 from transformers import Trainer, is_wandb_available
 
-from .utils import generate_model_card, get_comet_experiment_url
+from .utils import generate_model_card, get_comet_experiment_url, get_config_model_id
 
 
 if is_wandb_available():
@@ -50,8 +50,9 @@ def create_model_card(
         if not self.is_world_process_zero():
             return
 
-        if hasattr(self.model.config, "_name_or_path") and not os.path.isdir(self.model.config._name_or_path):
-            base_model = self.model.config._name_or_path
+        model_name_or_path = get_config_model_id(self.model.config)
+        if model_name_or_path and not os.path.isdir(model_name_or_path):
+            base_model = model_name_or_path
         else:
             base_model = None
 

trl/trainer/callbacks.py

@@ -40,7 +40,7 @@
 from ..mergekit_utils import MergeConfig, merge_models, upload_model_to_hf
 from ..models.utils import unwrap_model_for_generation
 from .judges import BasePairwiseJudge
-from .utils import log_table_to_comet_experiment
+from .utils import get_config_model_id, log_table_to_comet_experiment
 
 
 if is_rich_available():
@@ -821,7 +821,7 @@ def _merge_and_maybe_push(self, output_dir, global_step, model):
         checkpoint_path = os.path.join(output_dir, f"checkpoint-{global_step}")
         self.merge_config.policy_model_path = checkpoint_path
         if self.merge_config.target_model_path is None:
-            self.merge_config.target_model_path = model.config._name_or_path
+            self.merge_config.target_model_path = get_config_model_id(model.config)
         merge_path = os.path.join(checkpoint_path, "merged")
 
         merge_models(self.merge_config.create(), merge_path)

trl/trainer/dpo_trainer.py

@@ -65,6 +65,7 @@
     empty_cache,
     flush_left,
     flush_right,
+    get_config_model_id,
     log_table_to_comet_experiment,
     pad,
     pad_to_length,
@@ -286,7 +287,7 @@ def __init__(
     ):
         # Args
         if args is None:
-            model_name = model if isinstance(model, str) else model.config._name_or_path
+            model_name = model if isinstance(model, str) else get_config_model_id(model.config)
             model_name = model_name.split("/")[-1]
             args = DPOConfig(f"{model_name}-DPO")
 
@@ -299,7 +300,7 @@ def __init__(
                     "You passed `model_init_kwargs` to the `DPOConfig`, but your model is already instantiated. "
                     "The `model_init_kwargs` will be ignored."
                 )
-        model_id = model.config._name_or_path
+        model_id = get_config_model_id(model.config)
         if isinstance(ref_model, str):
             ref_model = create_model_from_path(ref_model, **args.ref_model_init_kwargs or {})
         else:

trl/trainer/grpo_trainer.py

@@ -66,6 +66,7 @@
     disable_dropout_in_model,
     ensure_master_addr_port,
     entropy_from_logits,
+    get_config_model_id,
     identity,
     nanmax,
     nanmin,
@@ -245,7 +246,7 @@ def __init__(
     ):
         # Args
         if args is None:
-            model_name = model if isinstance(model, str) else model.config._name_or_path
+            model_name = model if isinstance(model, str) else get_config_model_id(model.config)
             model_name = model_name.split("/")[-1]
             args = GRPOConfig(f"{model_name}-GRPO")
 
@@ -270,7 +271,7 @@ def __init__(
             architecture = getattr(transformers, config.architectures[0])
             model = architecture.from_pretrained(model_id, **model_init_kwargs)
         else:
-            model_id = model.config._name_or_path
+            model_id = get_config_model_id(model.config)
             if args.model_init_kwargs is not None:
                 logger.warning(
                     "You passed `model_init_kwargs` to the `GRPOConfig`, but your model is already instantiated. "
@@ -290,7 +291,7 @@ def __init__(
 
         # Processing class
         if processing_class is None:
-            processing_class = AutoProcessor.from_pretrained(model.config._name_or_path, truncation_side="left")
+            processing_class = AutoProcessor.from_pretrained(get_config_model_id(model.config), truncation_side="left")
 
         # Handle pad token for processors or tokenizers
         if isinstance(processing_class, ProcessorMixin):
@@ -317,7 +318,7 @@ def __init__(
                     reward_func, num_labels=1, **model_init_kwargs
                 )
             if isinstance(reward_funcs[i], nn.Module):  # Use Module over PretrainedModel for compat w/ compiled models
-                self.reward_func_names.append(reward_funcs[i].config._name_or_path.split("/")[-1])
+                self.reward_func_names.append(get_config_model_id(reward_funcs[i].config).split("/")[-1])
             else:
                 self.reward_func_names.append(reward_funcs[i].__name__)
         self.reward_funcs = reward_funcs
@@ -347,7 +348,7 @@ def __init__(
         for i, (reward_processing_class, reward_func) in enumerate(zip(reward_processing_classes, reward_funcs)):
             if isinstance(reward_func, PreTrainedModel):
                 if reward_processing_class is None:
-                    reward_processing_class = AutoTokenizer.from_pretrained(reward_func.config._name_or_path)
+                    reward_processing_class = AutoTokenizer.from_pretrained(get_config_model_id(reward_func.config))
                 if reward_processing_class.pad_token_id is None:
                     reward_processing_class.pad_token = reward_processing_class.eos_token
                 # The reward model computes the reward for the latest non-padded token in the input sequence.

trl/trainer/online_dpo_trainer.py

@@ -74,6 +74,7 @@
     disable_dropout_in_model,
     empty_cache,
     ensure_master_addr_port,
+    get_config_model_id,
     pad,
     truncate_right,
 )
@@ -243,7 +244,7 @@ def __init__(
                         reward_func, num_labels=1, **model_init_kwargs
                     )
                 if isinstance(reward_funcs[i], nn.Module):
-                    self.reward_func_names.append(reward_funcs[i].config._name_or_path.split("/")[-1])
+                    self.reward_func_names.append(get_config_model_id(reward_funcs[i].config).split("/")[-1])
                 else:
                     self.reward_func_names.append(reward_funcs[i].__name__)
             self.reward_funcs = reward_funcs

trl/trainer/reward_trainer.py

@@ -44,7 +44,7 @@
 from ..models import clone_chat_template, get_act_offloading_ctx_manager, prepare_peft_model
 from .base_trainer import BaseTrainer
 from .reward_config import RewardConfig
-from .utils import disable_dropout_in_model, pad, remove_none_values
+from .utils import disable_dropout_in_model, get_config_model_id, pad, remove_none_values
 
 
 if is_peft_available():
@@ -273,7 +273,7 @@ def __init__(
     ):
         # Args
         if args is None:
-            model_name = model if isinstance(model, str) else model.config._name_or_path
+            model_name = model if isinstance(model, str) else get_config_model_id(model.config)
             model_name = model_name.split("/")[-1]
             args = RewardConfig(f"{model_name}-Reward")
 
@@ -294,7 +294,7 @@ def __init__(
             with suppress_from_pretrained_warning(transformers.modeling_utils.logger):
                 model = AutoModelForSequenceClassification.from_pretrained(model_id, num_labels=1, **model_init_kwargs)
         else:
-            model_id = model.config._name_or_path
+            model_id = get_config_model_id(model.config)
             if args.model_init_kwargs is not None:
                 logger.warning(
                     "You passed `model_init_kwargs` to the `RewardConfig`, but your model is already instantiated. "

trl/trainer/rloo_trainer.py

@@ -65,6 +65,7 @@
     disable_dropout_in_model,
     ensure_master_addr_port,
     entropy_from_logits,
+    get_config_model_id,
     identity,
     nanmax,
     nanmin,
@@ -240,7 +241,7 @@ def __init__(
 
         # Args
         if args is None:
-            model_name = model if isinstance(model, str) else model.config._name_or_path
+            model_name = model if isinstance(model, str) else get_config_model_id(model.config)
             model_name = model_name.split("/")[-1]
             args = RLOOConfig(f"{model_name}-RLOO")
 
@@ -265,7 +266,7 @@ def __init__(
             architecture = getattr(transformers, config.architectures[0])
             model = architecture.from_pretrained(model_id, **model_init_kwargs)
         else:
-            model_id = model.config._name_or_path
+            model_id = get_config_model_id(model.config)
             if args.model_init_kwargs is not None:
                 logger.warning(
                     "You passed `model_init_kwargs` to the `RLOOConfig`, but your model is already instantiated. "
@@ -285,7 +286,7 @@ def __init__(
 
         # Processing class
         if processing_class is None:
-            processing_class = AutoProcessor.from_pretrained(model.config._name_or_path, truncation_side="left")
+            processing_class = AutoProcessor.from_pretrained(get_config_model_id(model.config), truncation_side="left")
 
         # Handle pad token for processors or tokenizers
         if isinstance(processing_class, ProcessorMixin):
@@ -312,7 +313,7 @@ def __init__(
                     reward_func, num_labels=1, **model_init_kwargs
                 )
             if isinstance(reward_funcs[i], nn.Module):  # Use Module over PretrainedModel for compat w/ compiled models
-                self.reward_func_names.append(reward_funcs[i].config._name_or_path.split("/")[-1])
+                self.reward_func_names.append(get_config_model_id(reward_funcs[i].config).split("/")[-1])
             else:
                 self.reward_func_names.append(reward_funcs[i].__name__)
         self.reward_funcs = reward_funcs
@@ -342,7 +343,7 @@ def __init__(
         for i, (reward_processing_class, reward_func) in enumerate(zip(reward_processing_classes, reward_funcs)):
             if isinstance(reward_func, PreTrainedModel):
                 if reward_processing_class is None:
-                    reward_processing_class = AutoTokenizer.from_pretrained(reward_func.config._name_or_path)
+                    reward_processing_class = AutoTokenizer.from_pretrained(get_config_model_id(reward_func.config))
                 if reward_processing_class.pad_token_id is None:
                     reward_processing_class.pad_token = reward_processing_class.eos_token
                 # The reward model computes the reward for the latest non-padded token in the input sequence.

trl/trainer/sft_trainer.py

@@ -54,6 +54,7 @@
     create_model_from_path,
     entropy_from_logits,
     flush_left,
+    get_config_model_id,
     pad,
     remove_none_values,
     selective_log_softmax,
@@ -590,7 +591,7 @@ def __init__(
     ):
         # Args
         if args is None:
-            model_name = model if isinstance(model, str) else model.config._name_or_path
+            model_name = model if isinstance(model, str) else get_config_model_id(model.config)
             model_name = model_name.split("/")[-1]
             args = SFTConfig(f"{model_name}-SFT")
         elif isinstance(args, TrainingArguments) and not isinstance(args, SFTConfig):
@@ -608,11 +609,10 @@ def __init__(
                     "You passed `model_init_kwargs` to the `SFTConfig`, but your model is already instantiated. "
                     "The `model_init_kwargs` will be ignored."
                 )
-        model_id = model.config._name_or_path
 
         # Processing class
         if processing_class is None:
-            processing_class = AutoProcessor.from_pretrained(model_id)
+            processing_class = AutoProcessor.from_pretrained(get_config_model_id(model.config))
 
         # Handle pad token for processors or tokenizers
         if isinstance(processing_class, ProcessorMixin):

trl/trainer/utils.py

@@ -41,6 +41,7 @@
     BitsAndBytesConfig,
     EvalPrediction,
     GenerationConfig,
+    PretrainedConfig,
     PreTrainedModel,
     PreTrainedTokenizerBase,
     TrainerState,
@@ -1962,3 +1963,19 @@ def create_model_from_path(model_id: str, **kwargs) -> PreTrainedModel:
     architecture = getattr(transformers, config.architectures[0])
     model = architecture.from_pretrained(model_id, **kwargs)
     return model
+
+
+def get_config_model_id(config: PretrainedConfig) -> str:
+    """
+    Retrieve the model identifier from a given model configuration.
+
+    Args:
+        config ([`~transformers.PreTrainedConfig`]):
+            Configuration from which to extract the model identifier.
+
+    Returns:
+        `str`:
+            The model identifier associated with the model configuration.
+    """
+    # Fall back to `config.text_config._name_or_path` if `config._name_or_path` is missing: Qwen2-VL and Qwen2.5-VL. See GH-4323
+    return getattr(config, "_name_or_path", "") or getattr(getattr(config, "text_config", None), "_name_or_path", "")