hf

app.py

@@ -7,7 +7,15 @@ import os
 
 login(token=os.environ["HUGGINGFACEHUB_API_TOKEN"])
 
-model = ChatHuggingFace(model="HuggingFaceH4/zephyr-7b-beta")
+llm = HuggingFaceEndpoint(
+    repo_id="HuggingFaceH4/zephyr-7b-beta",
+    task="text-generation",
+    max_new_tokens=512,
+    do_sample=False,
+    repetition_penalty=1.03,
+)
+
+model = ChatHuggingFace(llm=llm)
 
 def predict(message, history):
     history_langchain_format = []