Update app.py

app.py

@@ -166,7 +166,7 @@ class LLMService:
 
         self.sd_adapter.init_pipe(vae=vae,
                                   scheduler=noise_scheduler,
-                                  visual_encoder=self.visual_encoder.to("cpu"),
+                                  visual_encoder=self.visual_encoder.cpu(),
                                   image_transform=self.image_transform,
                                   discrete_model=None,
                                   dtype=self.dtype,
@@ -337,7 +337,7 @@ def generate(text_list, image_list, max_new_tokens, force_boi, force_bbox):
     if output['has_img_output']:
         print('loading visual encoder and llm to CPU, and sd to GPU')
         a = time.time()
-        service.agent = service.agent.to("cpu")
+        service.agent = service.agent.cpu()
         service.sd_adapter = service.sd_adapter.to(service.vit_sd_device, dtype=service.dtype)
         print("Loading finished: ", time.time() - a)
 
@@ -351,7 +351,7 @@ def generate(text_list, image_list, max_new_tokens, force_boi, force_bbox):
 
         print('loading visual encoder and llm to GPU, and sd to CPU')
         a = time.time()
-        service.sd_adapter = service.sd_adapter.to("cpu")
+        service.sd_adapter = service.sd_adapter.cpu()
         service.visual_encoder = service.visual_encoder.to(service.vit_sd_device, dtype=service.dtype)
         service.agent = service.agent.to(service.vit_sd_device, dtype=service.dtype)
         print("Loading finished: ", time.time() - a)