add llava-next/llava-next-video/video-llava

Former-commit-id: 6642cd501d

src/llamafactory/model/loader.py

@@ -119,15 +119,6 @@ def load_config(model_args: "ModelArguments") -> "PretrainedConfig":
     Loads model config.
     """
     init_kwargs = _get_init_kwargs(model_args)
-    if "LLaVA-NeXT-Video" in model_args.model_name_or_path:
-        from transformers import CLIPVisionConfig, LlamaConfig, LlavaNextVideoConfig, PretrainedConfig
-
-        official_config = PretrainedConfig.from_pretrained(model_args.model_name_or_path, **init_kwargs)
-        config = LlavaNextVideoConfig(
-            CLIPVisionConfig(**official_config.vision_config), LlamaConfig(**official_config.text_config)
-        )
-        setattr(config, "visual_inputs", True)
-        return config
     return AutoConfig.from_pretrained(model_args.model_name_or_path, **init_kwargs)
 
 
@@ -164,11 +155,6 @@ def load_model(
                 load_class = AutoModelForVision2Seq
             else:
                 load_class = AutoModelForCausalLM
-                if "llava_next_video" == getattr(config, "model_type"):
-                    from transformers import LlavaNextVideoForConditionalGeneration
-
-                    load_class = LlavaNextVideoForConditionalGeneration
-
             if model_args.train_from_scratch:
                 model = load_class.from_config(config)
             else:

src/llamafactory/model/model_utils/visual.py

@@ -92,7 +92,7 @@ def autocast_projector_dtype(model: "PreTrainedModel", model_args: "ModelArgumen
 
     if getattr(model, "quantization_method", None):
         model_type = getattr(model.config, "model_type", None)
-        if model_type in ["llava", "paligemma"]:
+        if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
             mm_projector: "torch.nn.Module" = getattr(model, "multi_modal_projector")
         elif model_type == "qwen2_vl":
             mm_projector: "torch.nn.Module" = getattr(getattr(model, "visual"), "merger")
@@ -111,9 +111,8 @@ def configure_visual_model(config: "PretrainedConfig") -> None:
     if model_type in [
         "llava",
         "llava_next",
-        "video_llava",
-        "idefics2",
         "llava_next_video",
+        "video_llava",
     ]:  # required for ds zero3 and valuehead models
         setattr(config, "hidden_size", getattr(config.text_config, "hidden_size", None))
 
@@ -128,7 +127,7 @@ def get_forbidden_modules(config: "PretrainedConfig", finetuning_args: "Finetuni
     """
     model_type = getattr(config, "model_type", None)
     forbidden_modules = set()
-    if model_type in ["llava", "paligemma"]:
+    if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
         if finetuning_args.freeze_vision_tower:
             forbidden_modules.add("vision_tower")
 
@@ -170,7 +169,7 @@ def patch_target_modules(
     """
     model_type = getattr(config, "model_type", None)
     if finetuning_args.freeze_vision_tower:
-        if model_type in ["llava", "paligemma"]:
+        if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
             return "^(?!.*vision_tower).*(?:{}).*".format("|".join(target_modules))
         elif model_type == "qwen2_vl":
             return "^(?!.*visual).*(?:{}).*".format("|".join(target_modules))