CosyVoice/runtime/triton_trtllm/model_repo_cosyvoice3/token2wav/config.pbtxt

name: "token2wav"
backend: "python"
max_batch_size: ${triton_max_batch_size}

dynamic_batching {
    max_queue_delay_microseconds: ${max_queue_delay_microseconds}
    priority_levels: 100
    default_priority_level: 100
}

parameters [
  {
   key: "model_dir",
   value: {string_value:"${model_dir}"}
  }
]

input [
  {
    name: "target_speech_tokens"
    data_type: TYPE_INT32
    dims: [-1]
  },
  {
    name: "prompt_speech_tokens"
    data_type: TYPE_INT32
    dims: [-1]
    optional: true
  },
  {
    name: "prompt_speech_feat"
    data_type: TYPE_FP16
    dims: [-1, 80]
    optional: true
  },
  {
    name: "prompt_spk_embedding"
    data_type: TYPE_FP16
    dims: [-1]
    optional: true
  },
  {
    name: "token_offset"
    data_type: TYPE_INT32
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  },
  {
    name: "finalize"
    data_type: TYPE_BOOL
    dims: [ 1 ]
    reshape: { shape: [ ] }
    optional: true
  }
]
output [
  {
    name: "mel"
    data_type: TYPE_FP32
    dims: [ 80, -1 ]
  }
]

instance_group [
  {
    count: 1
    kind: KIND_GPU
    gpus: [ 0 ]
  }
]