Update main.py

main.py

@@ -53,6 +53,7 @@ def generate_text(request: Request, input_data: InputData):
     top_p = input_data.top_p
     repetition_penalty = input_data.repetition_penalty
     input_text = generate_input_text(input_data)
+    max_new_tokens = min(max_new_tokens, 29500 - len(input_text))
     print(f"{datetime.now()} - Input Text: {input_text}")
     history = []
     generated_response = generate(input_text, history, temperature, max_new_tokens, top_p, repetition_penalty)