Update app.py

app.py

@@ -29,7 +29,7 @@ def query(payload):
     response = requests.request("POST", API_URL, json=payload)
     return json.loads(response.content.decode("utf-8"))
     
-def inference(input_sentence, max_length=16, temperature=0.1, greedy_decoding=False, top_k=0, top_p=1, seed=42):
+def inference(input_sentence, max_length, temperature, greedy_decoding, top_k, top_p, seed=42):
     top_k = None if top_k == 0 else top_k
     payload = {"inputs": input_sentence,
                "parameters": {"max_new_tokens": max_length, "top_k": top_k, "top_p": top_p, "temperature": temperature,
@@ -42,7 +42,13 @@ def inference(input_sentence, max_length=16, temperature=0.1, greedy_decoding=Fa
 
 gr.Interface(
     inference, 
-    [gr.inputs.Textbox(label="Input")], 
+    [
+        gr.inputs.Textbox(label="Input"),
+        gr.inputs.Slider(1, 64, value=8, label="Tokens to generate"),
+        gr.inputs.Slider(0, 64, value=0, label="Top K"),
+        gr.inputs.Slider(0.0, 10, value=0.9, step=0.05, label="Top P"),
+        gr.inputs.Checkbox(False, label="Greedy decoding"),
+    ], 
     gr.outputs.Textbox(label="Output"),
     examples=examples,
     # article=article,