# MOSS-TTS-Local-Transformer-v1.5 3-NPU deploy: Stage0 talker on NPU 0, Stage1 codec x2 replicas on NPU 1,2. # Splits AR (talker+depth) and codec across 3 devices, with 2 codec replicas for doubled codec throughput. # Launch: vllm-omni serve ... --deploy-config moss_tts_local_3npu.yaml base_config: moss_tts_local_2npu.yaml stages: - stage_id: 1 devices: "1,2" num_replicas: 2