[Feature] : Support disaggregated inference pipeline for Qwen3_TTS (#1161)

Signed-off-by: Sy03 <1370724210@qq.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

tests/worker/test_omni_gpu_model_runner.py

@@ -69,6 +69,7 @@ def _make_runner(req_ids=("r1", "r2"), hidden_size=4):
     runner.text_step = DummyBuffer(torch.zeros((bsz, hidden_size), dtype=torch.float32))
 
     runner.talker_mtp = DummyTalkerMTP()
+    runner.model = SimpleNamespace(talker_mtp_output_key="code_predictor_codes")
     runner.vllm_config = object()
 
     # Provide a minimal implementation that returns the expected 4-tuple.