down wan 14b

app.py

@@ -276,10 +276,12 @@ def ui(GPU_memory_mode, scheduler_dict, config_path, compile_dit, weight_dtype):
             # Use snapshot download for the VideoCoF repo to get all weights (including safetensors)
             try:
                 from huggingface_hub import snapshot_download
+                print("Downloading Wan2.1-T2V-14B weights...")
+                snapshot_download(repo_id="Wan-AI/Wan2.1-T2V-14B", local_dir="Wan-AI/Wan2.1-T2V-14B")
                 print("Downloading VideoCoF weights...")
                 snapshot_download(repo_id="XiangpengYang/VideoCoF")
             except Exception as e:
-                print(f"Warning: Failed to pre-download VideoCoF weights: {e}")
+                print(f"Warning: Failed to pre-download weights: {e}")
 
             base_model_dropdown, lora_model_dropdown, lora_alpha_slider, _ = create_finetune_models_checkpoints(controller, visible=False, default_lora="XiangpengYang/VideoCoF")
             

videox_fun/models/wan_image_encoder.py

@@ -537,8 +537,22 @@ class CLIPModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             return filtered_kwargs
 
         model = cls(**filter_kwargs(cls, transformer_additional_kwargs))
+        
+        # Handle HF hub download
+        import os
+        from huggingface_hub import hf_hub_download
+        
+        # If path doesn't exist locally, assume it's a repo ID and try to download
+        if not os.path.exists(pretrained_model_path):
+            try:
+                # Try to download models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth from the repo
+                print(f"Downloading CLIP model from {pretrained_model_path}...")
+                pretrained_model_path = hf_hub_download(repo_id=pretrained_model_path, filename="models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth")
+            except Exception as e:
+                print(f"Failed to download CLIP model from HF: {e}")
+        
         if pretrained_model_path.endswith(".safetensors"):
-            from safetensors.torch import load_file, safe_open
+            from safetensors.torch import load_file
             state_dict = load_file(pretrained_model_path)
         else:
             state_dict = torch.load(pretrained_model_path, map_location="cpu")
@@ -549,5 +563,5 @@ class CLIPModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         m, u = model.load_state_dict(state_dict)
         
         print(f"### missing keys: {len(m)}; \n### unexpected keys: {len(u)};")
-        print(m, u)
+        # print(m, u)
         return model

videox_fun/models/wan_text_encoder.py

@@ -311,6 +311,26 @@ class WanT5EncoderModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             valid_params = set(sig.parameters.keys()) - {'self', 'cls'}
             filtered_kwargs = {k: v for k, v in kwargs.items() if k in valid_params}
             return filtered_kwargs
+            
+        # Check if pretrained_model_path is a Hugging Face repo ID and download if needed
+        import os
+        from huggingface_hub import hf_hub_download
+        
+        # If path doesn't exist locally, assume it's a repo ID and try to download the weights
+        if not os.path.exists(pretrained_model_path):
+            try:
+                # Try to download models_t5_umt5-xxl-enc-bf16.pth from the repo
+                # Note: The user mentioned `models_t5_umt5-xxl-enc-bf16.pth` in previous context or similar.
+                # But here we should check the file name. Wan repo usually has `models_t5_umt5-xxl-enc-bf16.pth` inside a folder or root.
+                # However, usually we download the file that corresponds to this class.
+                # Let's assume the user passes the full path or we default to a standard name if it's a directory/repo.
+                # If pretrained_model_path is a repo ID, we need the filename.
+                # Based on `wan_civitai.yaml`: text_encoder_subpath: models_t5_umt5-xxl-enc-bf16.pth
+                print(f"Downloading text encoder from {pretrained_model_path}...")
+                pretrained_model_path = hf_hub_download(repo_id=pretrained_model_path, filename="models_t5_umt5-xxl-enc-bf16.pth")
+            except Exception as e:
+                print(f"Failed to download Text Encoder from HF: {e}")
+                # Fallback to original path logic which might fail later if file missing
     
         if low_cpu_mem_usage:
             try:

videox_fun/models/wan_vae.py

@@ -691,16 +691,32 @@ class AutoencoderKLWan(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             return filtered_kwargs
 
         model = cls(**filter_kwargs(cls, additional_kwargs))
+        
+        # Check if pretrained_model_path is a Hugging Face repo ID (e.g., "Wan-AI/Wan2.1-T2V-14B")
+        import os
+        from huggingface_hub import hf_hub_download
+        
+        # If path doesn't exist locally, assume it's a repo ID and try to download VAE file
+        if not os.path.exists(pretrained_model_path):
+            try:
+                # Try to download Wan2.1_VAE.pth from the repo
+                print(f"Downloading Wan2.1_VAE.pth from {pretrained_model_path}...")
+                pretrained_model_path = hf_hub_download(repo_id=pretrained_model_path, filename="Wan2.1_VAE.pth")
+            except Exception as e:
+                print(f"Failed to download VAE from HF: {e}")
+                # Fallback or re-raise if needed, but torch.load will fail anyway if path is invalid
+        
         if pretrained_model_path.endswith(".safetensors"):
-            from safetensors.torch import load_file, safe_open
+            from safetensors.torch import load_file
             state_dict = load_file(pretrained_model_path)
         else:
             state_dict = torch.load(pretrained_model_path, map_location="cpu")
+            
         tmp_state_dict = {} 
         for key in state_dict:
             tmp_state_dict["model." + key] = state_dict[key]
         state_dict = tmp_state_dict
         m, u = model.load_state_dict(state_dict, strict=False)
         print(f"### missing keys: {len(m)}; \n### unexpected keys: {len(u)};")
-        print(m, u)
+        # print(m, u)
         return model