Fix streaming decode with use_cuda_graph=True: first frame never runs, K-frame calls repeat the last frame

modeling_moss_audio_tokenizer.py

@@ -560,12 +560,15 @@ class MossAudioTokenizerDecodeSession:
             self._cuda_graph_key = cuda_graph_key
             self._graph_output_audio = decoder_output.audio
             self._graph_output_audio_lengths = decoder_output.audio_lengths
-        else:
-            self._cuda_graph.replay()
 
+        # Capture only records the kernels, so run the graph for this frame too.
+        self._cuda_graph.replay()
+
+        # Every replay overwrites these static buffers, while step() keeps each
+        # frame's output until the end of the call: hand out copies.
         return MossAudioTokenizerDecoderOutput(
-            audio=self._graph_output_audio,
-            audio_lengths=self._graph_output_audio_lengths,
+            audio=self._graph_output_audio.clone(),
+            audio_lengths=self._graph_output_audio_lengths.clone(),
         )
 
     def _reset_slot(self, slot_index: int) -> None: