diff --git a/GPT_SoVITS/AR/models/t2s_model.py b/GPT_SoVITS/AR/models/t2s_model.py
index 31c26cf..cecfcbc 100644
--- a/GPT_SoVITS/AR/models/t2s_model.py
+++ b/GPT_SoVITS/AR/models/t2s_model.py
@@ -579,8 +579,8 @@ class Text2SemanticDecoder(nn.Module):
 
             ####################### update next step ###################################
             y_emb = self.ar_audio_embedding(y[:, -1:])
-            xy_pos = y_emb * self.ar_audio_position.x_scale + self.ar_audio_position.alpha * self.ar_audio_position.pe[:, prompts.shape[1] + idx]
+            xy_pos = y_emb * self.ar_audio_position.x_scale + self.ar_audio_position.alpha * self.ar_audio_position.pe[:, y_len + idx].to(dtype=y_emb.dtype,device=y_emb.device)
 
         if ref_free:
             return y[:, :-1], 0
-        return y[:, :-1], idx - 1
+        return y[:, :-1], idx - 1
\ No newline at end of file