updated

app.py

@@ -6,11 +6,10 @@ DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 model = GPTLanguageModel().to(DEVICE)
 model.load_state_dict(torch.load("mini-gpt.pth",map_location=DEVICE), strict=False)
 model.eval()
-answer = decode(model.generate(context, max_new_tokens=500)[0].tolist())
 
 def display(text,number):
-    combined_text = text + answer[:number + 1]
-    return combined_text
+    answer = decode(model.generate(context, max_new_tokens=number)[0].tolist())
+    return answer
 
 input_box = gr.Textbox(label="Story Lines",value="Once Upon a Time")
 input_slider = gr.Slider(minimum=200, maximum=500, label="Select the maxium number of tokens/words:",step=100)