fix llava config

src/llmtuner/model/adapter.py

@@ -46,6 +46,9 @@ def init_adapter(
         if (not finetuning_args.pure_bf16) and (not finetuning_args.use_badam):
             model = model.float()
 
+        if model_args.visual_inputs and hasattr(model, "vision_tower"):  # freeze vision model
+            model.vision_tower.requires_grad_(False)
+
     if finetuning_args.finetuning_type == "freeze" and is_trainable:
         logger.info("Fine-tuning method: Freeze")
         num_layers = (

src/llmtuner/model/loader.py

@@ -106,7 +106,7 @@ def load_model(
     """
     init_kwargs = _get_init_kwargs(model_args)
     config = load_config(model_args)
-    patch_config(config, tokenizer, model_args, init_kwargs, is_trainable, add_valuehead)
+    patch_config(config, tokenizer, model_args, init_kwargs, is_trainable)
 
     model = None
     lazy_load = False

src/llmtuner/model/patcher.py

@@ -15,8 +15,8 @@
 from .utils.moe import add_z3_leaf_module, configure_moe
 from .utils.quantization import configure_quantization
 from .utils.rope import configure_rope
-from .utils.valuehead import configure_valuehead, prepare_valuehead_model
-from .utils.visual import autocast_projector_dtype
+from .utils.valuehead import prepare_valuehead_model
+from .utils.visual import autocast_projector_dtype, configure_hidden_size
 
 
 if TYPE_CHECKING:
@@ -40,7 +40,6 @@ def patch_config(
     model_args: "ModelArguments",
     init_kwargs: Dict[str, Any],
     is_trainable: bool,
-    add_valuehead: bool,
 ) -> None:
     if model_args.compute_dtype is None:  # priority: bf16 > fp16 > fp32
         model_args.compute_dtype = infer_optim_dtype(model_dtype=getattr(config, "torch_dtype", None))
@@ -50,9 +49,7 @@ def patch_config(
     configure_longlora(config, model_args, is_trainable)
     configure_quantization(config, tokenizer, model_args, init_kwargs)
     configure_moe(config, model_args, is_trainable)
-
-    if add_valuehead:
-        configure_valuehead(config)
+    configure_hidden_size(config)
 
     if model_args.use_cache and not is_trainable:
         setattr(config, "use_cache", True)

src/llmtuner/model/utils/valuehead.py

@@ -8,19 +8,14 @@
 
 
 if TYPE_CHECKING:
-    from transformers import PretrainedConfig, PreTrainedModel
+    from transformers import PreTrainedModel
 
     from ...hparams import ModelArguments
 
 
 logger = get_logger(__name__)
 
 
-def configure_valuehead(config: "PretrainedConfig") -> None:
-    if getattr(config, "model_type", None) == "llava":
-        setattr(config, "hidden_size", getattr(config.vision_config, "intermediate_size", None))
-
-
 def load_valuehead_params(path_or_repo_id: str, model_args: "ModelArguments") -> Dict[str, torch.Tensor]:
     r"""
     Loads value head parameters from Hugging Face Hub or local disk.

src/llmtuner/model/utils/visual.py

@@ -6,14 +6,19 @@
 
 
 if TYPE_CHECKING:
-    from transformers import PreTrainedModel
+    from transformers import PretrainedConfig, PreTrainedModel
 
     from ...hparams import ModelArguments
 
 
 logger = get_logger(__name__)
 
 
+def configure_hidden_size(config: "PretrainedConfig") -> None:
+    if getattr(config, "model_type", None) == "llava":
+        setattr(config, "hidden_size", getattr(config.text_config, "hidden_size", None))
+
+
 def autocast_projector_dtype(
     model: "PreTrainedModel", model_args: "ModelArguments", mm_projector_name: str = "multi_modal_projector"
 ) -> None:
@@ -22,7 +27,7 @@ def _mm_projector_forward_post_hook(
     ) -> "torch.Tensor":
         return output.to(model_args.compute_dtype)
 
-    if hasattr(model, mm_projector_name):
+    if hasattr(model, mm_projector_name) and getattr(model.config, "quantization_method", None):
         logger.info("Casting multimodal projector outputs in {}.".format(model_args.compute_dtype))
         mm_projector: "torch.nn.Module" = getattr(model, mm_projector_name)
         mm_projector.register_forward_hook(_mm_projector_forward_post_hook)