When max_batch_size=0, how to set cuda graph shape

[SunnyGhj]

Description
My question is: when max_batch_size=0, how to set cuda graph shape.

my model config

platform: "tensorrt_plan"
max_batch_size: 0
version_policy: { latest: { num_versions: 2}}
parameters { key: "execution_mode" value: { string_value: "0" } }
parameters { key: "intra_op_thread_count" value: { string_value: "4" } }
parameters { key: "inter_op_thread_count" value: { string_value: "4" } }
input [
  {
    name: "is_vip"
    data_type: TYPE_INT32
    dims: [-1, 1]
  },
  {
    name: "app_client_id"
    data_type: TYPE_INT32
    dims: [-1, 1]
  },
  {
    name: "keyword"
    data_type: TYPE_INT32
    dims: [-1, 1]
  },
  {
    name: "station_id"
    data_type: TYPE_INT32
    dims: [-1, 1]
  },
  {
    name: "city_number"
    data_type: TYPE_INT32
    dims: [-1, 1]
  },
  {
    name: "coreword_q"
    data_type: TYPE_INT32
    dims: [-1, 1]
  },
  {
    name: "aliasWordList"
    data_type: TYPE_INT32
    dims: [-1, 5]
  },
  {
    name: "extWordList"
    data_type: TYPE_INT32
    dims: [-1, 5]
  },
  {
    name: "cutWordList"
    data_type: TYPE_INT32
    dims: [-1, 10]
  },
  {
    name: "second_category_id"
    data_type: TYPE_INT32
    dims: [-1, 5]
  },
  {
    name: "third_category_id"
    data_type: TYPE_INT32
    dims: [-1, 5]
  },
  {
    name: "leaf_category_id"
    data_type: TYPE_INT32
    dims: [-1, 5]
  },
  {
    name: "PMID"
    data_type: TYPE_INT32
    dims: [-1, 1]
  },
  {
    name: "BRAND"
    data_type: TYPE_INT32
    dims: [-1, 1]
  }]
output [
  {
    name: "add_cvr"
    data_type: TYPE_FP16
    dims: [-1, 1]
  },
  {
    name: "order_cvr"
    data_type: TYPE_FP16
    dims: [-1, 1]
  }]
instance_group [ { count: 5 }]
optimization {
  graph: {level: 1}
  cuda: {
     graphs: true
     busy_wait_events: true
     graph_spec: [
      {
        batch_size: 0,
        input: { key: "is_vip", value: { dim: [1, 1] } },
        input: { key: "app_client_id", value: { dim: [1, 1] } },
        input: { key: "keyword", value: { dim: [1, 1] } },
        input: { key: "station_id", value: { dim: [1, 1] } },
        input: { key: "city_number", value: { dim: [1, 1] } },
        input: { key: "coreword_q", value: { dim: [1, 1] } },
        input: { key: "aliasWordList", value: { dim: [1, 5] } },
        input: { key: "extWordList", value: { dim: [1, 5] } },
        input: { key: "cutWordList", value: { dim: [1, 10] } },
        input: { key: "second_category_id", value: { dim: [1, 5] } },
        input: { key: "third_category_id", value: { dim: [1, 5] } },
        input: { key: "leaf_category_id", value: { dim: [1, 5] } },
        input: { key: "PMID", value: { dim: [200, 1] } }
      }
      ]
  }
  eager_batching: true
}

Set cuda graph seems to have no acceleration effect. Maybe there is something wrong with my model configuration. My question is, if max_batch_size=0, how should the input shape in the graph spec configuration be set?