Fix an issue in TTS generate to avoid gap between training and inference

modeling_minicpmo.py

@@ -2986,7 +2986,7 @@ class ConditionalChatTTS(PreTrainedModel):
                 inputs_embeds = torch.stack(code_emb, 3).sum(3)
 
             position_ids = torch.tensor(
-                [past_key_values[0][0].shape[2] + 1], dtype=torch.long, device=self.device
+                [past_key_values[0][0].shape[2]], dtype=torch.long, device=self.device
             ).unsqueeze(0)
 
             cache_position = position_ids.clone()