Update app.py

app.py

@@ -1,6 +1,4 @@
-SOTA_API_BASE="http://13.230.243.208:50001/v1"
-SOTA_API_KEY="sk-ceqvswstgcoviiwwfeuixbfloclfttdimjckjzamgavysnsn"
-SOTA_API_MODEL="Llama-3.1-8B--Instruct-syllabary"
+
 
 import gradio as gr
 import requests
@@ -26,12 +24,12 @@ def respond(
     response = ""
 
     headers = {
-        "Authorization": f"Bearer {SOTA_API_KEY}",
+        "Authorization": f"Bearer xxxx",
         "Content-Type": "application/json"
     }
 
     data = {
-        "model": SOTA_API_MODEL,
+        "model": "Llama-3.1-8B--Instruct-syllabary",
         "messages": messages,
         "max_tokens": max_tokens,
         "temperature": temperature,
@@ -39,7 +37,7 @@ def respond(
         "stream": True
     }
 
-    response = requests.post(f"{SOTA_API_BASE}/chat/completion", headers=headers, json=data)
+    response = requests.post("http://13.230.243.208:50001/v1/chat/completion", headers=headers, json=data)
 
     if response.status_code == 200:
         for token in response.iter_lines():