to(cuda)

lora.py

@@ -70,9 +70,9 @@ def submit_to_model():
     content = ""
     for text in text_streamer:
         print(text)
-        if text.endswith("<|eot_id|>"):
-            text = text.replace("<|eot_id|>", "")
         content += text
+        if content.endswith("<|eot_id|>"):
+            content = content.replace("<|eot_id|>", "")
         yield content