fix dtype

app.py

@@ -104,7 +104,6 @@ def get_response(params):
                 )
 
             images = [load_image_from_base64(image) for image in images]
-            print(images)
             # images = process_images(images, image_processor, model.config)
             images = process_images(images, image_processor, model.config).to('cpu', dtype=torch.float)
 
@@ -142,7 +141,7 @@ def get_response(params):
     input_ids = (
         tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt")
         .unsqueeze(0)
-        .to(model.device)
+        .to(model.device, dtype=torch.float)
     )
     keywords = [stop_str]
 
@@ -154,7 +153,7 @@ def get_response(params):
     max_new_tokens = min(
         max_new_tokens, max_context_length - input_ids.shape[-1] - num_image_tokens
     )
-
+    images = images.to(dtype=torch.float)
     if max_new_tokens < 1:
         yield json.dumps(
             {