logging

app.py

@@ -46,6 +46,7 @@ def generate(
     current_input += message
     
     device = "cuda"
+    print(current_input)
     input_ids = tokenizer(current_input, return_tensors="pt").input_ids.to(device)
     original_input_length = input_ids.shape[1]  # Remember the input length
 
@@ -68,11 +69,9 @@ def generate(
     )
     t = Thread(target=model.generate, kwargs=generate_kwargs)
     t.start()
-    print()
 
     outputs = []
     for text in streamer:
-        print(text, end="")
         outputs.append(text)
         yield "".join(outputs)