Advanced configuration

The default settings form TorchServe should be sufficient for most use cases. However, if you want to customize TorchServe, the configuration options described in this topic are available.

There are three ways to configure TorchServe. In order of priority, they are:

1. Environment variables

2. Command line arguments

3. Configuration file

For example, the value of an environment variable overrides both command line arguments and a property in the configuration file. The value of a command line argument overrides a value in the configuration file.

Environment variables

You can change TorchServe behavior by setting the following environment variables:

• JAVA_HOME

• PYTHONPATH

• TS_CONFIG_FILE

• LOG_LOCATION

• METRICS_LOCATION

Note: Environment variables have higher priority than command line or config.properties. The value of an environment variable overrides other property values.

Command line parameters

Customize TorchServe behavior by using the following command line arguments when you call torchserve:

• –ts-config TorchServe loads the specified configuration file if TS_CONFIG_FILE environment variable is not set

• –model-store Overrides the model_store property in config.properties file

• –models Overrides the load_models property in config.properties

• –log-config Overrides the default log4j2.xml

• –foreground Runs TorchServe in the foreground. If this option is disabled, TorchServe runs in the background

For more detailed information about torchserve command line options, see Serve Models with TorchServe.

config.properties file

TorchServe uses a config.properties file to store configurations. TorchServe uses following, in order of priority, to locate this config.properties file:

1. If the TS_CONFIG_FILE environment variable is set, TorchServe loads the configuration from the path specified by the environment variable.

2. If --ts-config parameter is passed to torchserve, TorchServe loads the configuration from the path specified by the parameter.

3. If there is a config.properties in the folder where you call torchserve, TorchServe loads the config.properties file from the current working directory.

4. If none of the above is specified, TorchServe loads a built-in configuration with default values.

Customize JVM options

To control TorchServe frontend memory footprint, configure the vmargs property in the config.properties file

• default: N/A, use JVM default options

Adjust JVM options to fit your memory requirement.

Load models at startup

You can configure TorchServe to load models during startup by setting the model_store and load_models properties. The following values are valid:

• load_models

  • standalone: default: N/A, No models are loaded at start up.

  • all: Load all models present in model_store.

  • model1.mar, model2.mar: Load models in the specified MAR files from model_store.

  • model1=model1.mar, model2=model2.mar: Load models with the specified names and MAR files from model_store.

• model_store

  • standalone: default: N/A, Loading models from the local disk is disabled.

  • pathname: The model store location is specified by the value of pathname.

Note: model_store and load_models properties are overridden by command line parameters, if specified.

Configure TorchServe listening address and port

TorchServe doesn’t support authentication natively. To avoid unauthorized access, TorchServe only allows localhost access by default. The inference API is listening on port 8080. The management API is listening on port 8081. Both expect HTTP requests. These are the default ports. See Enable SSL to configure HTTPS.

• inference_address: Inference API binding address. Default: http://127.0.0.1:8080

• management_address: Management API binding address. Default: http://127.0.0.1:8081

• metrics_address: Metrics API binding address. Default: http://127.0.0.1:8082

• To run predictions on models on a specific IP address, specify the IP address and port.

# bind inference API to localhost with SSL enabled
inference_address=https://127.0.0.1:8443

# bind inference API to private network interfaces with SSL enabled
inference_address=https://172.16.1.10:8080

Configure TorchServe gRPC listening addresses, ports and max connection age

The inference gRPC API is listening on port 7070, and the management gRPC API is listening on port 7071 on localhost by default.

To configure different addresses use following properties

• grpc_inference_address: Inference gRPC API IP address. Default: 127.0.0.1

• grpc_management_address: Management gRPC API IP address. Default: 127.0.0.1

To configure different ports use following properties

• grpc_inference_port: Inference gRPC API binding port. Default: 7070

• grpc_management_port: management gRPC API binding port. Default: 7071

To configure max connection age (milliseconds)

• grpc_inference_max_connection_age_ms: Inference gRPC max connection age. Default: Infinite

• grpc_management_max_connection_age_ms: Management gRPC max connection age. Default: Infinite

To configure max connection age grace (milliseconds)

• grpc_inference_max_connection_age_grace_ms: Inference gRPC max connection age grace. Default: Infinite

• grpc_management_max_connection_age_grace_ms: Management gRPC max connection age grace. Default: Infinite

Enable SSL

To enable HTTPs, you can change inference_address, management_address or metrics_address protocol from http to https. For example: inference_address=https://127.0.0.1. The default is port 443, but you can make TorchServe listen on whatever port you set to accept https requests. For example, to receive https traffic on port 8443, you would use: inference_address=https://127.0.0.1:8443.

You must also provide a certificate and private key to enable SSL. TorchServe supports two ways to configure SSL:

1. Use a keystore:

• keystore: the keystore file location. If multiple private key entries exist in the keystore, the first one will be used.

• keystore_pass: the keystore password. The password (if applicable) MUST be the same as keystore password.

• keystore_type: the type of keystore. Default: PKCS12.

1. Use private-key/certificate files:

• private_key_file: the private key file location. Supports both PKCS8 and OpenSSL private keys.

• certificate_file: the X509 certificate chain file location.

Examples

Option 1: Use a keystore; generate a keystore with Java’s keytool. Note the storepass argument expects you to create your own password.

keytool -genkey -keyalg RSA -alias ts -keystore keystore.p12 -storepass changeit -storetype PKCS12 -validity 3600 -keysize 2048 -dname "CN=www.MY_TS.com, OU=Cloud Service, O=model server, L=Palo Alto, ST=California, C=US"

Configure the following properties in config.properties:

inference_address=https://127.0.0.1:8443
management_address=https://127.0.0.1:8444
metrics_address=https://127.0.0.1:8445
keystore=keystore.p12
keystore_pass=changeit
keystore_type=PKCS12

Option 2: Use private-key/certificate files; generate your self signed cert and key with OpenSSL:

openssl req -x509 -nodes -days 365 -newkey rsa:2048 -keyout mykey.key -out mycert.pem

Config following property in config.properties:

inference_address=https://127.0.0.1:8443
management_address=https://127.0.0.1:8444
metrics_address=https://127.0.0.1:8445
private_key_file=mykey.key
certificate_file=mycert.pem

Configure Cross-Origin Resource Sharing (CORS)

CORS is a mechanism that uses additional HTTP headers to tell a browser to let a web application running at one origin (domain) have permission to access selected resources from a server at a different origin.

CORS is disabled by default. Configure following properties in config.properties file to enable CORS:

# cors_allowed_origin is required to enable CORS, use '*' or your domain name
cors_allowed_origin=https://yourdomain.com
# required if you want to use preflight request
cors_allowed_methods=GET, POST, PUT, OPTIONS
# required if the request has an Access-Control-Request-Headers header
cors_allowed_headers=X-Custom-Header

Prefer direct buffer

Configuration parameter prefer_direct_buffer controls if the model server will be using direct memory specified by -XX:MaxDirectMemorySize. This parameter is for model server only and doesn’t affect other packages’ usage of direct memory buffer. Default: false

prefer_direct_buffer=true

Allow model specific custom python packages.

Custom models/handlers may depend on different python packages which are not installed by-default as a part of TorchServe setup. User can supply a requirements.txt file containing the required list of python packages to be installed by TorchServe for seamless model serving. Configuration parameter install_py_dep_per_model controls if the model server will install the python packages using the requirements file supplied with the model archive. Default: false

install_py_dep_per_model=true

User can also supply custom python packages in zip or tar.gz format using the --extra-files flag while creating the model-archive and make an entry of the file name in the requirements file.

Restrict backend worker to access environment variables

Environment variables might contain sensitive information, like AWS credentials. Backend workers execute an arbitrary model’s custom code, which might expose a security risk. TorchServe provides a blacklist_env_vars property that allows you to restrict which environment variables can be accessed by backend workers.

• blacklist_env_vars: a regular expression to filter out environment variable names. Default: all environment variables are visible to backend workers.

Limit GPU usage

By default, TorchServe uses all available GPUs for inference. Use number_of_gpu to limit the usage of GPUs.

• number_of_gpu: Maximum number of GPUs that TorchServe can use for inference. Default: all available GPUs in system.

Nvidia control Visibility

Set nvidia environment variables. For example:

• export CUDA_DEVICE_ORDER=”PCI_BUS_ID”

• export CUDA_VISIBLE_DEVICES=”1,3”

Enable metrics api

• enable_metrics_api : Enable or disable metric apis i.e. it can be either true or false. Default: true (Enabled)

Config model

• models: Use this to set configurations specific to a model. The value is presented in json format.

{
    "modelName": {
        "version": {
            "parameterName1": parameterValue1,
            "parameterName2": parameterValue2,
            "parameterNameN": parameterValueN,
        }
    }
}

A model’s parameters are defined in model source code

• minWorkers: the minimum number of workers of a model

• maxWorkers: the maximum number of workers of a model

• batchSize: the batch size of a model

• maxBatchDelay: the maximum delay in msec of a batch of a model

• responseTimeout: the timeout in sec of a specific model’s response. This setting takes priority over default_response_timeout which is a default timeout over all models

• defaultVersion: the default version of a model

• marName: the mar file name of a model

A model’s configuration example

models={\
  "noop": {\
    "1.0": {\
        "defaultVersion": true,\
        "marName": "noop.mar",\
        "minWorkers": 1,\
        "maxWorkers": 1,\
        "batchSize": 4,\
        "maxBatchDelay": 100,\
        "responseTimeout": 120\
    }\
  },\
  "vgg16": {\
    "1.0": {\
        "defaultVersion": true,\
        "marName": "vgg16.mar",\
        "minWorkers": 1,\
        "maxWorkers": 4,\
        "batchSize": 8,\
        "maxBatchDelay": 100,\
        "responseTimeout": 120\
    }\
  }\
}

Starting from version 0.8.0, TorchServe allows for model configuration using a YAML file embedded in the MAR file. This YAML file contains two distinct parts that determine how a model is configured: frontend parameters and backend parameters. (see details)

• The frontend parameters are controlled by TorchServe’s frontend and specify the parameter name and default values. TorchServe now uses a priority order to determine the final value of a model’s parameters in frontend. Specifically, the config.property file has the lowest priority, followed by the model configuration YAML file, and finally, the REST or gRPC model management API has the highest priority.

• The backend parameters are fully controlled by the user. Users customized handler can access the backend parameters via the model_yaml_config property of the context object. For example, context.model_yaml_config[”pippy”][”rpc_timeout”].

• User can allocate specific GPU device IDs to a model by defining “deviceIds” in the frontend parameters in the YAML file. TorchServe uses a round-robin strategy to assign device IDs to a model’s worker. If specified in the YAML file, it round-robins the device IDs listed; otherwise, it uses all visible device IDs on the host.

Other properties

Most of the following properties are designed for performance tuning. Adjusting these numbers will impact scalability and throughput.

• enable_envvars_config: Enable configuring TorchServe through environment variables. When this option is set to “true”, all the static configurations of TorchServe can come through environment variables as well. Default: false

• number_of_netty_threads: Number frontend netty thread. This specifies the number of threads in the child EventLoopGroup of the frontend netty server. This group provides EventLoops for processing Netty Channel events (namely inference and management requests) from accepted connections. Default: number of logical processors available to the JVM.

• netty_client_threads: Number of backend netty thread. This specifies the number of threads in the WorkerThread EventLoopGroup which writes inference responses to the frontend. Default: number of logical processors available to the JVM.

• default_workers_per_model: Number of workers to create for each model that loaded at startup time. Default: available GPUs in system or number of logical processors available to the JVM.

• job_queue_size: Number inference jobs that frontend will queue before backend can serve. Default: 100.

• async_logging: Enable asynchronous logging for higher throughput, log output may be delayed if this is enabled. Default: false.

• default_response_timeout: Timeout, in seconds, used for all models backend workers before they are deemed unresponsive and rebooted. Default: 120 seconds.

• unregister_model_timeout: Timeout, in seconds, used when handling an unregister model request when cleaning a process before it is deemed unresponsive and an error response is sent. Default: 120 seconds.

• decode_input_request: Configuration to let backend workers to decode requests, when the content type is known. If this is set to “true”, backend workers do “Bytearray to JSON object” conversion when the content type is “application/json” and the backend workers convert “Bytearray to utf-8 string” when the Content-Type of the request is set to “text*”. Default: true

• initial_worker_port : This is the initial port number for auto assigning port to worker process.

• model_store : Path of model store directory.

• model_server_home : Torchserve home directory.

• max_request_size : The maximum allowable request size that the Torchserve accepts, in bytes. Default: 6553500

• max_response_size : The maximum allowable response size that the Torchserve sends, in bytes. Default: 6553500

• limit_max_image_pixels : Default value is true (Use default PIL.Image.MAX_IMAGE_PIXELS). If this is set to “false”, set PIL.Image.MAX_IMAGE_PIXELS = None in backend default vision handler for large image payload.

• allowed_urls : Comma separated regex of allowed source URL(s) from where models can be registered. Default: file://.*|http(s)?://.* (all URLs and local file system) e.g. : To allow base URLs https://s3.amazonaws.com/ and https://torchserve.pytorch.org/ use the following regex string allowed_urls=https://s3.amazonaws.com/.*,https://torchserve.pytorch.org/.*

  • For security reason, use_env_allowed_urls=true is required in config.properties to read allowed_urls from environment variable.

• workflow_store : Path of workflow store directory. Defaults to model store directory.

• disable_system_metrics : Disable collection of system metrics when set to “true”. Default value is “false”.

• system_metrics_cmd: The customized system metrics python script name with arguments. For example:ts/metrics/metric_collector.py --gpu 0. Default: empty which means TorchServe collects system metrics via “ts/metrics/metric_collector.py –gpu $CUDA_VISIBLE_DEVICES”.

NOTE

All the above config properties can be set using environment variable as follows.

• set enable_envvars_config to true in config.properties

• export environment variable for property asTS_<PROPERTY_NAME>.

  e.g.: to set inference_address property run cmd export TS_INFERENCE_ADDRESS="http://127.0.0.1:8082".

---