Update app.py

app.py

@@ -11,6 +11,9 @@ import random
 from openai import OpenAI
 import subprocess
 
+LLAMA_3B_API_ENDPOINT = os.environ.get("LLAMA_3B_API_ENDPOINT")
+LLAMA_3B_API_KEY = os.environ.get("LLAMA_3B_API_KEY")
+
 default_lang = "en"
 
 engines = { default_lang: Model(default_lang) }
@@ -28,8 +31,8 @@ HF_TOKEN = os.environ.get("HF_TOKEN", None)
 def client_fn(model):
     if "Llama 3 8B Service" in model:
         return OpenAI(
-            base_url="http://52.76.81.56:60002/v1",
-            api_key="token-abc123"
+            base_url=LLAMA_3B_API_ENDPOINT,
+            api_key=LLAMA_3B_API_KEY
         )
     elif "Llama" in model:
         return InferenceClient("meta-llama/Meta-Llama-3-8B-Instruct")