Update app.py

app.py

@@ -114,7 +114,8 @@ async def tts(ws: WebSocket):
                 max_new_tokens  = CHUNK_TOKENS,
                 logits_processor= [masker],
                 do_sample=True, temperature=0.7, top_p=0.95,
-                use_cache=True
+                use_cache=True,
+                return_legacy_cache=True # Added legacy cache
             )
 
             # ----- neue Tokens heraus schneiden --------------------------