Skip to content

Initialization

The following page provides more details about possible options for configuring the Triton Inference Server and working with block and non-blocking mode for tests and deployment.

Configuring Triton

Connecting Python models with Triton Inference Server working in the current environment requires creating a Triton object. This can be done by creating a context:

from pytriton.triton import Triton

with Triton() as triton:
    ...

or simply creating an object:

from pytriton.triton import Triton

triton = Triton()

The Triton Inference Server behavior can be configured by passing config parameter:

import pathlib
from pytriton.triton import Triton, TritonConfig

triton_config = TritonConfig(log_file=pathlib.Path("/tmp/triton.log"))
with Triton(config=triton_config) as triton:
    ...

and through environment variables, for example, set as in the command below:

PYTRITON_TRITON_CONFIG_LOG_VERBOSITY=4 python my_script.py

The order of precedence of configuration methods is:

  • config defined through config parameter of Triton class __init__ method
  • config defined in environment variables
  • default TritonConfig values

Blocking mode

The blocking mode will stop the execution of the current thread and wait for incoming HTTP/gRPC requests for inference execution. This mode makes your application behave as a pure server. The example of using blocking mode:

from pytriton.triton import Triton

with Triton() as triton:
    ...  # Load models here
    triton.serve()

Background mode

The background mode runs Triton as a subprocess and does not block the execution of the current thread. In this mode, you can run Triton Inference Server and interact with it from the current context. The example of using background mode:

from pytriton.triton import Triton

triton = Triton()
...  # Load models here
triton.run()  # Triton Server started
print("This print will appear")
triton.stop()  # Triton Server stopped

Filesystem usage

PyTriton needs to access the filesystem for two purposes:

  • to communicate with the Triton backend using file sockets,
  • storing copy of Triton backend and its binary dependencies.

PyTriton creates temporary folders called Workspaces, where it stores the file descriptors for these operations. By default, these folders are located in $HOME/.cache/pytriton directory. However, you can change this location by setting the PYTRITON_HOME environment variable.