Update app.py

app.py

@@ -130,7 +130,7 @@ pipe.to(device=DEVICE)
 from transformers import AutoProcessor, PaliGemmaForConditionalGeneration, BitsAndBytesConfig
 
 quantization_config = BitsAndBytesConfig(load_in_4bit=True)
-pali = PaliGemmaForConditionalGeneration.from_pretrained('google/paligemma-3b-pt-224', torch_dtype=dtype,).eval().to('cuda')
+pali = PaliGemmaForConditionalGeneration.from_pretrained('google/paligemma-3b-pt-224', torch_dtype=dtype, device='cuda').eval()
 processor = AutoProcessor.from_pretrained('google/paligemma-3b-pt-224')
 
 
@@ -393,7 +393,7 @@ def start(_, calibrate_prompts, user_id, request: gr.Request):
             im = torch.nn.functional.interpolate(im, (224, 224))
             im = (im - .5) * 2
         
-            im_emb, gemb = encode_space(image, im.to('cuda'))
+            im_emb, gemb = encode_space(image, im)
             im_emb = im_emb.to('cpu')
             gemb = gemb.to('cpu')
 
@@ -573,7 +573,7 @@ scheduler.start()
 #thread.start()
 
 # TODO shouldn't call this before gradio launch, yeah?
-@spaces.GPU()
+@spaces.GPU(duration=20)
 def encode_space(x, im):
     with torch.no_grad():
         print('encode')