Update app.py

app.py

@@ -59,11 +59,11 @@ def get_model(temperature, top_p, repetition_penalty):
            "temperature": temperature,
            "top_p": top_p,
            "repetition_penalty": repetition_penalty,
-           "max_length": 2000
+           "max_length": 1000
        },
        huggingfacehub_api_token=huggingface_token
    )
-def generate_chunked_response(model, prompt, max_tokens=2000, max_chunks=5):
+def generate_chunked_response(model, prompt, max_tokens=1000, max_chunks=5):
    full_response = ""
    for i in range(max_chunks):
        chunk = model(prompt + full_response, max_new_tokens=max_tokens)