This document describes Triton's model repository extension. The model-repository extension allows a client to query and control the one or more model repositories being served by Triton. Because this extension is supported, Triton reports “model_repository” in the extensions field of the Server Metadata. This extension has an optional component, described below, that allows the unload API to specify the "unload_dependents" parameter. Versions of Triton that support this optional component will also report "model_repository(unload_dependents)" in the extensions field of the Server Metadata.
In all JSON schemas shown in this document $number, $string, $boolean,
$object and $array refer to the fundamental JSON types. #optional
indicates an optional JSON field.
The model-repository extension requires Index, Load and Unload APIs. Triton exposes the endpoints at the following URLs.
POST v2/repository/index
POST v2/repository/models/${MODEL_NAME}/load
POST v2/repository/models/${MODEL_NAME}/unload
The index API returns information about every model available in a model repository, even if it is not currently loaded into Triton. The index API provides a way to determine which models can potentially be loaded by the Load API. A model-repository index request is made with an HTTP POST to the index endpoint. In the corresponding response the HTTP body contains the JSON response.
The index request object, identified as $repository_index_request, is
required in the HTTP body of the POST request.
$repository_index_request =
{
"ready" : $boolean #optional,
}
- "ready" : Optional, default is false. If true return only models ready for inferencing.
A successful index request is indicated by a 200 HTTP status code. The
response object, identified as $repository_index_response, is returned
in the HTTP body for every successful request.
$repository_index_response =
[
{
"name" : $string,
"version" : $string #optional,
"state" : $string,
"reason" : $string
},
…
]
- “name” : The name of the model.
- “version” : The version of the model.
- “state” : The state of the model.
- “reason” : The reason, if any, that the model is in the current state.
A failed index request must be indicated by an HTTP error status
(typically 400). The HTTP body must contain the
$repository_index_error_response object.
$repository_index_error_response =
{
"error": $string
}
- “error” : The descriptive message for the error.
The load API requests that a model be loaded into Triton, or reloaded
if the model is already loaded. A load request is made with an HTTP
POST to a load endpoint. The HTTP body may be empty or may contain
the load request object, identified as $repository_load_request.
A successful load request is indicated by a 200 HTTP status.
$repository_load_request =
{
"parameters" : $parameters #optional
}
- "parameters" : An object containing zero or more parameters for this request expressed as key/value pairs. See Parameters for more information.
The load API accepts the following parameters:
-
"config" : string parameter that contains a JSON representation of the model configuration, which must be able to be parsed into ModelConfig message from model_config.proto. This config will be used for loading the model instead of the one in the model directory. If config is provided, the (re-)load will be triggered as the model metadata has been updated, and the same (re-)load behavior will be applied.
-
"file:<version>/<file-name>" : The serialized model file, base64 encoded. This convention will be used to specify the override model directory to load the model from. For instance, if the user wants to specify a model directory that contains an ONNX model as version 2, then the user will specify the parameter to "file:2/model.onnx" : "". Note that "config" parameter must be provided to serve as the model configuration of the override model directory.
A failed load request must be indicated by an HTTP error status
(typically 400). The HTTP body must contain the
$repository_load_error_response object.
$repository_load_error_response =
{
"error": $string
}
- “error” : The descriptive message for the error.
For the following request, Triton will load the model "mymodel" with provided model configuration and model file.
POST /v2/repository/models/mymodel/load HTTP/1.1
Host: localhost:8000
{
"parameters": {
"config": "{
"name": "mymodel",
"backend": "onnxruntime",
"inputs": [{
"name": "INPUT0",
"datatype": "FP32",
"shape": [ 1 ]
}
],
"outputs": [{
"name": "OUTPUT0",
"datatype": "FP32",
"shape": [ 1 ]
}
]
}",
"file:1/model.onnx" : "<base64-encoded-file-content>"
}
}
The unload API requests that a model be unloaded from Triton. An
unload request is made with an HTTP POST to an unload endpoint. The
HTTP body may be empty or may contain the unload request object,
identified as $repository_unload_request. A successful unload request
is indicated by a 200 HTTP status.
$repository_unload_request =
{
"parameters" : $parameters #optional
}
- "parameters" : An object containing zero or more parameters for this request expressed as key/value pairs. See Parameters for more information.
The unload API accepts the following parameters:
- "unload_dependents" : boolean parameter indicating that in addition to unloading the requested model, also unload any dependent model that was loaded along with the requested model. For example, request to unload the models composing an ensemble will unload the ensemble as well.
A failed unload request must be indicated by an HTTP error status
(typically 400). The HTTP body must contain the
$repository_unload_error_response object.
$repository_unload_error_response =
{
"error": $string
}
- “error” : The descriptive message for the error.
The model-repository extension requires the following API:
service GRPCInferenceService
{
…
// Get the index of model repository contents.
rpc RepositoryIndex(RepositoryIndexRequest)
returns (RepositoryIndexResponse) {}
// Load or reload a model from a repository.
rpc RepositoryModelLoad(RepositoryModeLoadRequest)
returns (RepositoryModelLoadResponse) {}
// Unload a model.
rpc RepositoryModelUnload(RepositoryModelUnloadRequest)
returns (RepositoryModelUnloadResponse) {}
}
message ModelRepositoryParameter
{
// The parameter value can be a string, an int64, a boolean
// or a message specific to a predefined parameter.
oneof parameter_choice
{
// A boolean parameter value.
bool bool_param = 1;
// An int64 parameter value.
int64 int64_param = 2;
// A string parameter value.
string string_param = 3;
// A bytes parameter value.
bytes bytes_param = 4;
}
}
The RepositoryIndex API returns information about every model available in a model repository, even if it is not currently loaded into Triton. Errors are indicated by the google.rpc.Status returned for the request. The OK code indicates success and other codes indicate failure. The request and response messages for RepositoryIndex are:
message RepositoryIndexRequest
{
// The name of the repository. If empty the index is returned
// for all repositories.
string repository_name = 1;
// If true return only models currently ready for inferencing.
bool ready = 2;
}
message RepositoryIndexResponse
{
// Index entry for a model.
message ModelIndex {
// The name of the model.
string name = 1;
// The version of the model.
string version = 2;
// The state of the model.
string state = 3;
// The reason, if any, that the model is in the given state.
string reason = 4;
}
// An index entry for each model.
repeated ModelIndex models = 1;
}
The RepositoryModelLoad API requests that a model be loaded into Triton, or reloaded if the model is already loaded. Errors are indicated by the google.rpc.Status returned for the request. The OK code indicates success and other codes indicate failure. The request and response messages for RepositoryModelLoad are:
message RepositoryModelLoadRequest
{
// The name of the repository to load from. If empty the model
// is loaded from any repository.
string repository_name = 1;
// The name of the model to load, or reload.
string model_name = 2;
// Optional parameters.
map<string, ModelRepositoryParameter> parameters = 3;
}
message RepositoryModelLoadResponse
{
}
The RepositoryModelLoad API accepts the following parameters:
-
"config" : string parameter that contains a JSON representation of the model configuration, which must be able to be parsed into ModelConfig message from model_config.proto. This config will be used for loading the model instead of the one in the model directory. If config is provided, the (re-)load will be triggered as the model metadata has been updated, and the same (re-)load behavior will be applied.
-
"file:<version>/<file-name>" : bytes parameter that contains the model file content. This convention will be used to specify the override model directory to load the model from. For instance, if the user wants to specify a model directory that contains an ONNX model as version 2, then the user will specify the parameter to "file:2/model.onnx" : "". Note that "config" parameter must be provided to serve as the model configuration of the override model directory.
The RepositoryModelUnload API requests that a model be unloaded from Triton. Errors are indicated by the google.rpc.Status returned for the request. The OK code indicates success and other codes indicate failure. The request and response messages for RepositoryModelUnload are:
message RepositoryModelUnloadRequest
{
// The name of the repository from which the model was originally
// loaded. If empty the repository is not considered.
string repository_name = 1;
// The name of the model to unload.
string model_name = 2;
// Optional parameters.
map<string, ModelRepositoryParameter> parameters = 3;
}
message RepositoryModelUnloadResponse
{
}
The RepositoryModelUnload API accepts the following parameters:
- "unload_dependents" : boolean parameter indicating that in addition to unloading the requested model, also unload any dependent model that was loaded along with the requested model. For example, request to unload the models composing an ensemble will unload the ensemble as well.