[Bugfix][Qwen3-TTS] Read full payload from model buffer

Signed-off-by: Sy03 <1370724210@qq.com>

tests/model_executor/models/qwen3_tts/test_qwen3_tts_code2wav.py

@@ -25,6 +25,7 @@ class _FakeDecoder(nn.Module):
         self.total_upsample = total_upsample
         self.decode_calls: list[dict[str, object]] = []
         self.batched_decode_calls: list[dict[str, object]] = []
+        self.batched_decode_codes: list[torch.Tensor] = []
         self.decode_codes: list[torch.Tensor] = []
         self.cudagraph_calls: list[dict[str, int | torch.device]] = []
 
@@ -64,6 +65,7 @@ class _FakeDecoder(nn.Module):
         left_context_size: int = 25,
         max_batch_size: int = 0,
     ) -> list[torch.Tensor]:
+        self.batched_decode_codes.append(codes.detach().cpu().clone())
         self.batched_decode_calls.append(
             {
                 "chunk_size": chunk_size,
@@ -203,6 +205,26 @@ def test_forward_uses_decoder_audio_contract_without_context():
     torch.testing.assert_close(audio, expected)
 
 
+def test_forward_reads_current_model_intermediate_buffer_for_full_payload():
+    """Full-payload sync mode must decode connector codec ids, not placeholders."""
+    model = _make_model()
+    placeholder_ids = torch.zeros(12, dtype=torch.long)
+    payload_codes = torch.arange(12, dtype=torch.long) + 17
+
+    model.forward(
+        input_ids=placeholder_ids,
+        runtime_additional_information=[],
+        model_intermediate_buffer=[
+            {
+                "codes": {"audio": payload_codes},
+                "meta": {"left_context_size": 0},
+            }
+        ],
+    )
+
+    torch.testing.assert_close(model.decoder.batched_decode_codes[-1], payload_codes.reshape(1, _NUM_QUANTIZERS, 6))
+
+
 @pytest.mark.parametrize("skipped_length", [0, 1], ids=["empty", "malformed"])
 def test_stateless_batch_uses_original_request_index_for_left_context(skipped_length):
     model = _make_model()

vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_code2wav.py

@@ -260,6 +260,7 @@ class Qwen3TTSCode2Wav(nn.Module):
         intermediate_tensors: Any = None,
         inputs_embeds: torch.Tensor | None = None,
         runtime_additional_information: list[dict[str, Any]] | None = None,
+        model_intermediate_buffer: list[dict[str, Any]] | None = None,
         **kwargs: Any,
     ) -> OmniOutput:
         """Decode codec codes into audio waveform.
@@ -285,6 +286,13 @@ class Qwen3TTSCode2Wav(nn.Module):
                 multimodal_outputs={"model_outputs": [empty], "sr": [sr_tensor]},
             )
 
+        # vLLM's runner renamed this per-request side channel to
+        # ``model_intermediate_buffer``.  Keep accepting the old explicit
+        # argument for older runners, and use the new name when it is empty.
+        # Without this fallback, the non-async-chunk/full-payload path decodes
+        # placeholder input_ids instead of the codec payload from Stage 0.
+        if not runtime_additional_information:
+            runtime_additional_information = model_intermediate_buffer
         runtime_infos = runtime_additional_information or []
         ids = input_ids.reshape(-1).to(dtype=torch.long)
         request_ids_list = self._split_request_ids(ids, kwargs.get("seq_token_counts"))