Update app.py

app.py

@@ -24,7 +24,7 @@ MAX_NEW_TOKENS = 25
 GPU_LAYERS = 0
 STOP_LIST=["###","##"]
 
-stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=64)])
+#stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=64)])
 
 from huggingface_hub import hf_hub_download
 hf_hub_download(repo_id="gorkemgoknar/llama2-7f-moviechatbot-ggml-q4", local_dir=".", filename="llama2-7f-fp16-ggml-q4.bin")
@@ -87,7 +87,7 @@ def get_audio_url(text,character):
     
 def get_response_cpp(prompt):
 
-    output = llm(prompt, max_tokens=32, stop=["#","sierpeda"], echo=True,stopping_criteria=stopping_criteria)
+    output = llm(prompt, max_tokens=32, stop=["#","sierpeda"], echo=True)
     #print(output)
     response_text= output["choices"][0]["text"]