Skip to content

Installing Narwhal from PyPI

narwhal-inference installs the narwhal Python package and six commands for local development, engine launch, attestation, profiling, preflight and routing. The package requires Linux and Python 3.11 or newer.

Install the package in a virtual environment:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install narwhal-inference
narwhal-serve --version
narwhal --help

Verifying the installation

Expected narwhal-serve --version output:

narwhal-inference <version>
  • narwhal --help lists the six installed commands.
  • Each command accepts --version before its operational arguments.
  • Each --version invocation exits with status 0.

For each deployment:

  1. Record the version with the fleet configuration, engine image, and profiles.
  2. Pin the approved version on every router host.

Before serving requests

Qualify these production inputs by following Deploying a fleet:

  • separately provisioned vLLM engines with NIXL, or SGLang engines with Mooncake
  • engine attestation
  • a fleet configuration
  • measured profiles

Installing from a source checkout

In a repository checkout, make setup installs Narwhal and its development dependencies into .venv.

If --version prints narwhal-inference unknown (distribution metadata unavailable), run python -m pip install -e ..