update

app.py

@@ -12,7 +12,7 @@ model = AutoModelForCausalLM.from_pretrained(
 )
 
 
-def predict(prompt): 
+def predict(prompt, return_thinking_content=False): 
     
     # prepare the model input
     messages = [
@@ -45,8 +45,10 @@ def predict(prompt):
     
     # print("thinking content:", thinking_content)
     # print("content:", content)
-
-    return thinking_content, content
+    if return_thinking_content:
+        return thinking_content, content
+    else:
+        return content
 
 
 demo = gr.Interface(fn=predict, inputs="text", outputs="text")

test.py

@@ -0,0 +1,13 @@
+from gradio_client import Client
+
+client = Client("simeonjunker/modeltest")
+print(client)
+
+prompt = 'Give me a short introduction to large language models.'
+
+result = client.predict(
+	prompt=prompt,
+	api_name="/predict"
+)
+
+print(result)