Fix vision_tower ZeroGPU compatibility.

model_builder.py

@@ -140,9 +140,9 @@ def load_pretrained_model(model_path, model_base, model_name, load_8bit=False, l
         if not vision_tower.is_loaded:
             vision_tower.load_model()
         if load_bf16:
-            vision_tower.to(device='cuda', dtype=torch.bfloat16)
+            vision_tower.to(device='cpu', dtype=torch.bfloat16)
         else:
-            vision_tower.to(device='cuda', dtype=torch.float16)
+            vision_tower.to(device='cpu', dtype=torch.float16)
         image_processor = vision_tower.image_processor
 
     if hasattr(model.config, "max_sequence_length"):

model_worker.py

@@ -52,11 +52,11 @@ class ModelWorker:
                 torch_device='cpu',
                 device_map="cpu",
             )
-        self.model.to('cuda')
 
     @spaces.GPU
     def generate_stream(self, params):
         tokenizer, model, image_processor = self.tokenizer, self.model, self.image_processor
+        model.to('cuda')
         logger.info(f'Model devices: {model.device}')
 
         prompt = params["prompt"]