Update app.py

app.py

@@ -153,7 +153,7 @@ LLM_STOP_WORDS= ["</s>","<|user|>","/s>"]
 LLAMA_VERBOSE=False
 
 print("Running LLM Zephyr")
-llm_zephyr = Llama(model_path=zephyr_model_path,n_gpu_layers=GPU_LAYERS-10,max_new_tokens=256, context_window=4096, n_ctx=4096,n_batch=128,verbose=LLAMA_VERBOSE)
+llm_zephyr = Llama(model_path=zephyr_model_path,n_gpu_layers=GPU_LAYERS-10,max_new_tokens=512, context_window=4096, n_ctx=4096,n_batch=128,verbose=LLAMA_VERBOSE)
 
 def split_sentences(text, max_len):
     # Apply custom rules to enforce sentence breaks with double punctuation