infer code

app.py

@@ -146,18 +146,21 @@ def render_context(file):
 
 
 @spaces.GPU(duration=120)
+def infer(context, query):
+    return model.query_longcite(
+        context=context,
+        query=query,
+        tokenizer=tokenizer,
+        max_input_length=128000,
+        max_new_tokens=1024,
+    )
+
 def run_llm(context, query):
     if not context:
         raise gr.Error("Error: no uploaded document")
     if not query:
         raise gr.Error("Error: no query")
-    result = model.query_longcite(
-        context,
-        query,
-        tokenizer=tokenizer,
-        max_input_length=128000,
-        max_new_tokens=1024,
-    )
+    result = infer(context=context, query=query)
     all_statements = result["all_statements"]
     answer_html, all_cite_html, clicked_cite_html, cite_num2idx_dict = convert_to_html(
         all_statements