Update app.py

app.py

@@ -136,6 +136,7 @@ async def tts(ws: WebSocket):
                     do_sample=True, temperature=0.7, top_p=0.95,
                     use_cache=True,
                     return_dict_in_generate=True,
+                    cache_position=torch.tensor([offset_len], device=device) # Explicitly pass cache_position
                 )
             print(f"DEBUG: After generate - type of gen.past_key_values: {type(gen.past_key_values)}", flush=True) # Added logging