cpu or cuda

app.py

@@ -529,15 +529,15 @@ def main():
     # model, quantizer, tokenizer, width, height, device = construction_layout()
     width, height, device = construction_layout()
     
-    inference_partial = partial(
-        inference,
-        model=model,
-        quantizer=quantizer,
-        tokenizer=tokenizer,
-        width=width,
-        height=height,
-        device=device
-    )
+    # inference_partial = partial(
+    #     inference,
+    #     model=model,
+    #     quantizer=quantizer,
+    #     tokenizer=tokenizer,
+    #     width=width,
+    #     height=height,
+    #     device=device
+    # )
     
     def process_preddate(intention, temperature, top_p, generate_method='v1'):
         intention = intention.replace('\n', '').replace('\r', '').replace('\\', '')