bugfix

app.py

@@ -108,10 +108,6 @@ def chat(
 ):
     global model
     history = history or []
-
-    if len(history) == 0:
-        # no history, so lets reset chat state
-        model.resetState()
         
     if model == None:
         gc.collect()
@@ -119,6 +115,10 @@ def chat(
             torch.cuda.empty_cache()
         model = get_model()
         
+    if len(history) == 0:
+        # no history, so lets reset chat state
+        model.resetState()
+        
     max_new_tokens = int(max_new_tokens)
     temperature = float(temperature)
     top_p = float(top_p)