69 lines
939 B
Plaintext
69 lines
939 B
Plaintext
# Triton backend to use
|
|
backend: "python"
|
|
max_batch_size: 0
|
|
|
|
# Triton should expect as input a single string
|
|
# input of variable length named 'text_input'
|
|
input [
|
|
|
|
{
|
|
name: "text_input"
|
|
data_type: TYPE_STRING
|
|
dims: [ -1 ]
|
|
|
|
|
|
},
|
|
{
|
|
name: "stream"
|
|
data_type: TYPE_BOOL
|
|
dims: [ 1 ]
|
|
|
|
optional: true
|
|
|
|
|
|
}
|
|
|
|
]
|
|
|
|
|
|
# Triton should expect to respond with a single string
|
|
# output of variable length named 'text_output'
|
|
output [
|
|
|
|
{
|
|
name: "text_output"
|
|
data_type: TYPE_STRING
|
|
dims: [ -1 ]
|
|
|
|
}
|
|
|
|
]
|
|
|
|
parameters: [
|
|
{
|
|
key: "base_model_path",
|
|
value: {string_value: "/cheetah/input/model/groupuser/triton-python-model"}
|
|
},
|
|
{
|
|
key: "is_adapter_model",
|
|
value: {string_value: "false"}
|
|
},
|
|
{
|
|
key: "adapter_model_path",
|
|
value: {string_value: ""}
|
|
},
|
|
|
|
{
|
|
key: "quantization",
|
|
value: {string_value: "none"}
|
|
}
|
|
]
|
|
|
|
instance_group [
|
|
{
|
|
kind: KIND_AUTO
|
|
count: 1
|
|
}
|
|
]
|
|
|