OMLE Server

Inference server for OMLE models, speaking the Open Inference Protocol over both REST and gRPC.

pip install omle-server
omle-server                        # built-in defaults
omle-server --help                 # every flag
omle-server --model-dir ./models --rest-port 9000
omle-server --config /etc/omle/server.json

This package is a delivery vehicle for a compiled binary, not a Python library. Installing it puts omle-server on PATH; that console script execs the native executable, so no Python remains in the process once the server is running.

Requirements

OpenSSL 3 must be present on the system. Everything else the server needs — gRPC, Abseil, protobuf, Drogon, jsoncpp, simdjson and the OMLE runtime itself — is linked statically into the binary, so there is nothing else to install.

OpenSSL is the deliberate exception. Bundling it would freeze a cryptographic library at build time, cut off from the security updates the operating system provides, and it could then only be patched by releasing a new wheel. It ships with every current Linux distribution and is in the Homebrew baseline on macOS:

# only if it is somehow missing
apt install libssl3          # Debian / Ubuntu
dnf install openssl-libs     # Fedora / RHEL
brew install openssl@3       # macOS

Windows is the exception: there is no system OpenSSL to link, so that wheel carries its own statically. Nothing to install, but a fix there needs a new release rather than an operating-system update.

Wheels are published for linux-x86_64, macos-arm64 and windows-x86_64. Other platforms build from source — see the repository README.

Configuring

Four sources, each overriding the one above it: built-in defaults, then a JSON config file, then the environment, then command-line flags.

omle-server init-config -o server.json    # every setting, at its default
omle-server --config server.json

OMLE_MODEL_DIR=./models omle-server --rest-port 9000
omle-server --help                        # the full list

Environment sits below flags so a container image can set a baseline that docker run still overrides.

There is deliberately no Python wrapper translating arguments into these. It could only restate what the binary already does, and would drift from it. To launch the server from Python, run it as a subprocess like any other program:

import os, subprocess
import omle_server

proc = subprocess.Popen(
    [omle_server.binary_path()],
    env={**os.environ, "OMLE_MODEL_DIR": "/models"},
)

omle_server.binary_path() returns the absolute path to the bundled executable, and is the only function this package exposes.

Checking it works

curl localhost:8080/v2/health/ready          # 200 once the registry has loaded
curl localhost:8080/v2/models/<name>         # input/output metadata

Related packages

• omle — the model IR, protobuf I/O and validation
• omle-convert — converters from trained scikit-learn, XGBoost, LightGBM, CatBoost and Spark ML models
• omle-runtime — the C++ inference runtime, with a scikit-learn-style API
• omle-spark — Spark ML transformer for scoring DataFrames from PySpark
• omle-viewer — interactive DAG viewer for Jupyter and the browser

License

Apache-2.0