Skip to content

Latest commit

 

History

History
552 lines (440 loc) · 20.1 KB

management_api.md

File metadata and controls

552 lines (440 loc) · 20.1 KB
Raw

Management API

TorchServe provides the following APIs that allows you to manage models at runtime:

  1. Register a model
  2. Increase/decrease number of workers for specific model
  3. Describe a model's status
  4. Unregister a model
  5. List registered models
  6. Set default version of a model
  7. Refresh tokens for token authorization

The Management API listens on port 8081 and is only accessible from localhost by default. To change the default setting, see TorchServe Configuration.

Management API for registering and deleting models is disabled by default. Add --enable-model-api to command line when running TorchServe to enable the use of these APIs. For more details and ways to enable see Model API control

For all Management API requests, TorchServe requires the correct Management token to be included or token authorization must be disabled. For more details see token authorization documentation

Similar to the Inference API, the Management API provides a API description to describe management APIs with the OpenAPI 3.0 specification.

Alternatively, if you want to use KServe, TorchServe supports both v1 and v2 API. For more details please look into this kserve documentation

Register a model

This API follows the ManagementAPIsService.RegisterModel gRPC API.

To use this API after TorchServe starts, model API control has to be enabled. Add --enable-model-api to command line when starting TorchServe to enable the use of this API. For more details see model API control

POST /models

  • url - Model archive download url. Supports the following locations:
    • a local model archive (.mar); the file must be in the model_store folder (and not in a subfolder).
    • a URI using the HTTP(s) protocol. TorchServe can download .mar files from the Internet.
  • model_name - the name of the model; this name will be used as {model_name} in other APIs as part of the path. If this parameter is not present, modelName in MANIFEST.json will be used.
  • handler - the inference handler entry-point. This value will override handler in MANIFEST.json if present. NOTE: Make sure that the given handler is in the PYTHONPATH. The format of handler is module_name:method_name.
  • runtime - the runtime for the model custom service code. This value will override runtime in MANIFEST.json if present. The default value is PYTHON.
  • batch_size - the inference batch size. The default value is 1.
  • max_batch_delay - the maximum delay for batch aggregation. The default value is 100 milliseconds.
  • initial_workers - the number of initial workers to create. The default value is 0. TorchServe will not run inference until there is at least one work assigned.
  • synchronous - whether or not the creation of worker is synchronous. The default value is false. TorchServe will create new workers without waiting for acknowledgement that the previous worker is online.
  • response_timeout - If the model's backend worker doesn't respond with inference response within this timeout period, the worker will be deemed unresponsive and rebooted. The units is seconds. The default value is 120 seconds.
  • startup_timeout - If the model's backend worker doesn't load the model within this timeout period, the worker will be deemed unresponsive and rebooted. The units is seconds. The default value is 120 seconds.
curl -X POST  "http://localhost:8081/models?url=https://torchserve.pytorch.org/mar_files/squeezenet1_1.mar"

{
  "status": "Model \"squeezenet_v1.1\" Version: 1.0 registered with 0 initial workers. Use scale workers API to add workers for the model."
}

Encrypted model serving

If you'd like to serve an encrypted model then you need to setup S3 SSE-KMS with the following environment variables:

  • AWS_ACCESS_KEY_ID
  • AWS_SECRET_ACCESS_KEY
  • AWS_DEFAULT_REGION

And set "s3_sse_kms=true" in HTTP request.

For example: model squeezenet1_1 is encrypted on S3 under your own private account. The model http url on S3 is https://torchserve.pytorch.org/sse-test/squeezenet1_1.mar.

  • if torchserve will run on EC2 instance (e.g. OS: ubuntu)
  1. add an IAM Role (AWSS3ReadOnlyAccess) for the EC2 instance
  2. run ts_scripts/get_aws_credential.sh to export AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY
  3. export AWS_DEFAULT_REGION=your_s3_bucket_region
  4. start torchserve
  5. Register encrypted model squeezenet1_1 by setting s3_sse_kms=true in curl command.
curl -X POST  "http://localhost:8081/models?url=https://torchserve.pytorch.org/sse-test/squeezenet1_1.mar&s3_sse_kms=true"

{
  "status": "Model \"squeezenet_v1.1\" Version: 1.0 registered with 0 initial workers. Use scale workers API to add workers for the model."
}
  • if torchserve will run on local (e.g. OS: macOS)
  1. Find your AWS access key and secret key. You can reset them if you forgot the keys.
  2. export AWS_ACCESS_KEY_ID=your_aws_access_key
  3. export AWS_SECRET_ACCESS_KEY=your_aws_secret_key
  4. export AWS_DEFAULT_REGION=your_s3_bucket_region
  5. start torchserve
  6. Register encrypted model squeezenet1_1 by setting s3_sse_kms=true in curl command (same as EC2 example step 5).

You might want to create workers during registration. because creating initial workers might take some time, you can choose between synchronous or asynchronous call to make sure initial workers are created properly.

The asynchronous call returns with HTTP code 202 before trying to create workers.

curl -v -X POST "http://localhost:8081/models?initial_workers=1&synchronous=false&url=https://torchserve.pytorch.org/mar_files/squeezenet1_1.mar"

< HTTP/1.1 202 Accepted
< content-type: application/json
< x-request-id: 4dc54158-c6de-42aa-b5dd-ebcb5f721043
< content-length: 47
< connection: keep-alive
<
{
  "status": "Processing worker updates..."
}

The synchronous call returns with HTTP code 200 after all workers have been adjusted.

curl -v -X POST "http://localhost:8081/models?initial_workers=1&synchronous=true&url=https://torchserve.pytorch.org/mar_files/squeezenet1_1.mar"

< HTTP/1.1 200 OK
< content-type: application/json
< x-request-id: ecd2e502-382f-4c3b-b425-519fbf6d3b85
< content-length: 89
< connection: keep-alive
<
{
  "status": "Model \"squeezenet1_1\" Version: 1.0 registered with 1 initial workers"
}

Scale workers

This API follows the ManagementAPIsService.ScaleWorker gRPC API. It returns the status of a model in the ModelServer.

PUT /models/{model_name}

  • min_worker - (optional) the minimum number of worker processes. TorchServe will try to maintain this minimum for specified model. The default value is 1.
  • max_worker - (optional) the maximum number of worker processes. TorchServe will make no more that this number of workers for the specified model. The default is the same as the setting for min_worker.
  • synchronous - whether or not the call is synchronous. The default value is false.
  • timeout - the specified wait time for a worker to complete all pending requests. If exceeded, the work process will be terminated. Use 0 to terminate the backend worker process immediately. Use -1 to wait infinitely. The default value is -1.

Use the Scale Worker API to dynamically adjust the number of workers for any version of a model to better serve different inference request loads.

There are two different flavors of this API, synchronous and asynchronous.

The asynchronous call will return immediately with HTTP code 202:

curl -v -X PUT "http://localhost:8081/models/noop?min_worker=3"

< HTTP/1.1 202 Accepted
< content-type: application/json
< x-request-id: 42adc58e-6956-4198-ad07-db6c620c4c1e
< content-length: 47
< connection: keep-alive
<
{
  "status": "Processing worker updates..."
}

The synchronous call returns with HTTP code 200 after all workers have been adjusted.

curl -v -X PUT "http://localhost:8081/models/noop?min_worker=3&synchronous=true"

< HTTP/1.1 200 OK
< content-type: application/json
< x-request-id: b72b1ea0-81c6-4cce-92c4-530d3cfe5d4a
< content-length: 63
< connection: keep-alive
<
{
  "status": "Workers scaled to 3 for model: noop"
}

To scale workers of a specific version of a model use URI : /models/{model_name}/{version} PUT /models/{model_name}/{version}

The following synchronous call will return after all workers for version "2.0" for model "noop" has be adjusted with HTTP code 200.

curl -v -X PUT "http://localhost:8081/models/noop/2.0?min_worker=3&synchronous=true"

< HTTP/1.1 200 OK
< content-type: application/json
< x-request-id: 3997ccd4-ae44-4570-b249-e361b08d3d47
< content-length: 77
< connection: keep-alive
<
{
  "status": "Workers scaled to 3 for model: noop, version: 2.0"
}

Describe model

This API follows the ManagementAPIsService.DescribeModel gRPC API. It returns the status of a model in the ModelServer.

GET /models/{model_name}

Use the Describe Model API to get detail runtime status of default version of a model:

curl http://localhost:8081/models/noop
[
    {
      "modelName": "noop",
      "modelVersion": "1.0",
      "modelUrl": "noop.mar",
      "engine": "Torch",
      "runtime": "python",
      "minWorkers": 1,
      "maxWorkers": 1,
      "batchSize": 1,
      "maxBatchDelay": 100,
      "workers": [
        {
          "id": "9000",
          "startTime": "2018-10-02T13:44:53.034Z",
          "status": "READY",
          "gpu": false,
          "memoryUsage": 89247744
        }
      ],
      "jobQueueStatus": {
        "remainingCapacity": 100,
        "pendingRequests": 0
      }
    }
]

GET /models/{model_name}/{version}

Use the Describe Model API to get detail runtime status of specific version of a model:

curl http://localhost:8081/models/noop/2.0
[
    {
      "modelName": "noop",
      "modelVersion": "2.0",
      "modelUrl": "noop_2.mar",
      "engine": "Torch",
      "runtime": "python",
      "minWorkers": 1,
      "maxWorkers": 1,
      "batchSize": 1,
      "maxBatchDelay": 100,
      "workers": [
        {
          "id": "9000",
          "startTime": "2018-10-02T13:44:53.034Z",
          "status": "READY",
          "gpu": false,
          "memoryUsage": 89247744
        }
      ],
      "jobQueueStatus": {
        "remainingCapacity": 100,
        "pendingRequests": 0
      }
    }
]

GET /models/{model_name}/all

Use the Describe Model API to get detail runtime status of all version of a model:

curl http://localhost:8081/models/noop/all
[
    {
      "modelName": "noop",
      "modelVersion": "1.0",
      "modelUrl": "noop.mar",
      "engine": "Torch",
      "runtime": "python",
      "minWorkers": 1,
      "maxWorkers": 1,
      "batchSize": 1,
      "maxBatchDelay": 100,
      "workers": [
        {
          "id": "9000",
          "startTime": "2018-10-02T13:44:53.034Z",
          "status": "READY",
          "gpu": false,
          "memoryUsage": 89247744
        }
      ],
      "jobQueueStatus": {
        "remainingCapacity": 100,
        "pendingRequests": 0
      }
    },
    {
      "modelName": "noop",
      "modelVersion": "2.0",
      "modelUrl": "noop_2.mar",
      "engine": "Torch",
      "runtime": "python",
      "minWorkers": 1,
      "maxWorkers": 1,
      "batchSize": 1,
      "maxBatchDelay": 100,
      "workers": [
        {
          "id": "9000",
          "startTime": "2018-10-02T13:44:53.034Z",
          "status": "READY",
          "gpu": false,
          "memoryUsage": 89247744
        }
      ],
      "jobQueueStatus": {
        "remainingCapacity": 100,
        "pendingRequests": 0
      }
    }
]

GET /models/{model_name}/{model_version}?customized=true or GET /models/{model_name}?customized=true

Use the Describe Model API to get detail runtime status and customized metadata of a version of a model:

  • Implement function describe_handle. E.g.
    def describe_handle(self):
        """Customized describe handler
        Returns:
            dict : A dictionary response.
        """
        output_describe = None

        logger.info("Collect customized metadata")

        return output_describe
  • Implement function _is_describe if handler is not inherited from BaseHandler. And then, call _is_describe and describe_handle in handle.
    def _is_describe(self):
        if self.context and self.context.get_request_header(0, "describe"):
            if self.context.get_request_header(0, "describe") == "True":
                return True
        return False

    def handle(self, data, context):
        if self._is_describe():
            output = [self.describe_handle()]
        else:
            data_preprocess = self.preprocess(data)

            if not self._is_explain():
                output = self.inference(data_preprocess)
                output = self.postprocess(output)
            else:
                output = self.explain_handle(data_preprocess, data)

        return output
  • Call function _is_describe and describe_handle in handle. E.g.
def handle(self, data, context):
        """Entry point for default handler. It takes the data from the input request and returns
           the predicted outcome for the input.
        Args:
            data (list): The input data that needs to be made a prediction request on.
            context (Context): It is a JSON Object containing information pertaining to
                               the model artifacts parameters.
        Returns:
            list : Returns a list of dictionary with the predicted response.
        """

        # It can be used for pre or post processing if needed as additional request
        # information is available in context
        start_time = time.time()

        self.context = context
        metrics = self.context.metrics

        is_profiler_enabled = os.environ.get("ENABLE_TORCH_PROFILER", None)
        if is_profiler_enabled:
            output, _ = self._infer_with_profiler(data=data)
        else:
            if self._is_describe():
                output = [self.describe_handle()]
            else:
                data_preprocess = self.preprocess(data)

                if not self._is_explain():
                    output = self.inference(data_preprocess)
                    output = self.postprocess(output)
                else:
                    output = self.explain_handle(data_preprocess, data)

        stop_time = time.time()
        metrics.add_time('HandlerTime', round(
            (stop_time - start_time) * 1000, 2), None, 'ms')
        return output
  • Here is an example. "customizedMetadata" shows the metadata from user's model. These metadata can be decoded into a dictionary.
curl http://localhost:8081/models/noop-customized/1.0?customized=true
[
    {
        "modelName": "noop-customized",
        "modelVersion": "1.0",
        "modelUrl": "noop-customized.mar",
        "runtime": "python",
        "minWorkers": 1,
        "maxWorkers": 1,
        "batchSize": 1,
        "maxBatchDelay": 100,
        "loadedAtStartup": false,
        "workers": [
          {
            "id": "9010",
            "startTime": "2022-02-08T11:03:20.974Z",
            "status": "READY",
            "memoryUsage": 0,
            "pid": 98972,
            "gpu": false,
            "gpuUsage": "N/A"
          }
        ],
        "jobQueueStatus": {
          "remainingCapacity": 100,
          "pendingRequests": 0
        },
        "customizedMetadata": "{\n  \"data1\": \"1\",\n  \"data2\": \"2\"\n}"
     }
]
  • Decode customizedMetadata on client side. For example:
import requests
import json

response = requests.get('http://localhost:8081/models/noop-customized/?customized=true').json()
customizedMetadata = response[0]['customizedMetadata']
print(customizedMetadata)

Unregister a model

This API follows the ManagementAPIsService.UnregisterModel gRPC API. It returns the status of a model in the ModelServer.

To use this API after TorchServe starts, model API control has to be enabled. Add --enable-model-api to command line when starting TorchServe to enable the use of this API. For more details see model API control

DELETE /models/{model_name}/{version}

Use the Unregister Model API to free up system resources by unregistering specific version of a model from TorchServe:

curl -X DELETE http://localhost:8081/models/noop/1.0

{
  "status": "Model \"noop\" unregistered"
}

List models

This API follows the ManagementAPIsService.ListModels gRPC API. It returns the status of a model in the ModelServer.

GET /models

  • limit - (optional) the maximum number of items to return. It is passed as a query parameter. The default value is 100.
  • next_page_token - (optional) queries for next page. It is passed as a query parameter. This value is return by a previous API call.

Use the Models API to query default versions of current registered models:

curl "http://localhost:8081/models"

This API supports pagination:

curl "http://localhost:8081/models?limit=2&next_page_token=2"

{
  "nextPageToken": "4",
  "models": [
    {
      "modelName": "noop",
      "modelUrl": "noop-v1.0"
    },
    {
      "modelName": "noop_v0.1",
      "modelUrl": "noop-v0.1"
    }
  ]
}

API Description

OPTIONS /

To view a full list of inference and management APIs, you can use following command:

# To view all inference APIs:
curl -X OPTIONS http://localhost:8080

# To view all management APIs:
curl -X OPTIONS http://localhost:8081

The out is OpenAPI 3.0.1 json format. You use it to generate client code, see swagger codegen for detail.

Example outputs of the Inference and Management APIs:

Set Default Version

This API follows the ManagementAPIsService.SetDefault gRPC API. It returns the status of a model in the ModelServer.

PUT /models/{model_name}/{version}/set-default

To set any registered version of a model as default version use:

curl -v -X PUT http://localhost:8081/models/noop/2.0/set-default

The out is OpenAPI 3.0.1 json format. You use it to generate client code, see swagger codegen for detail.

Token Authorization API

TorchServe now enforces token authorization by default. Check the following documentation for more information: Token Authorization.

This API is used in order to generate a new key to replace either the management or inference key.

Management Example:

curl localhost:8081/token?type=management -H "Authorization: Bearer {API Token}"

will replace the current management key in the key_file with a new one and will update the expiration time.

Inference example:

curl localhost:8081/token?type=inference -H "Authorization: Bearer {API Token}"

will replace the current inference key in the key_file with a new one and will update the expiration time.