On a laptop, dragonfarm builds its own Python environment the first time it runs. A server behind a firewall can’t, and a team often wants every machine to be exactly the same. For both, dragonfarm ships a lock: every Python package it needs, pinned to one version, with hashes, for Windows, macOS and Linux. This page is for the person who sets machines up.
The environment a laptop builds for itself uses the same pinned
versions, so it matches a server’s and does not change when new releases
appear. Set DRAGONFARM_PYTHON_LOCKED=false for the newest
releases instead.
You need uv and Python 3.12 or later on each machine. The lock comes with the package:
system.file("python", "requirements.lock", package = "dragonfarm")Same packages everywhere
On a machine with the internet, build the environment from the lock:
uv venv /opt/dragonfarm-py --python 3.12
uv pip sync requirements.lock --python /opt/dragonfarm-py --require-hashesThen tell dragonfarm to use it, for example in the server’s
.Renviron:
(On Windows the interpreter is Scripts\python.exe inside
the folder.)
No internet at all
On a machine that has the internet and the same operating system as the target, download the packages into a folder:
uv venv dl --python 3.12 --seed
dl/bin/python -m pip download -r requirements.lock --require-hashes --only-binary=:all: -d wheelsCopy wheels and requirements.lock across,
then build with no index:
uv venv /opt/dragonfarm-py --python 3.12
uv pip sync requirements.lock --python /opt/dragonfarm-py --require-hashes --offline --no-index --find-links wheelsOn Windows this was about 210 MB of packages. Built this way with an empty cache and no network, the environment came out identical to one built online.
Models come from Hugging Face. Download them on a connected machine,
copy the Hugging Face cache folder across, and set
HF_HUB_OFFLINE=1. Presets are pinned to exact model
versions, so every machine loads the same weights.
Check a machine
When DRAGONFARM_PYTHON is set, the check compares the
environment with the lock and names any package that differs.
GPUs
The lock is resolved from PyPI. On Windows that gives the CPU build of torch; on Linux, PyPI’s torch already includes CUDA. For a Windows server with an NVIDIA card, add the CUDA build from the PyTorch package index when you download.
One catalogue for the team
Each runs folder keeps a small record of what happened: answers and
where they came from, approvals, policies, model versions and who signed
them off. On a laptop that is a file in the folder. For a team, point
everyone at one Postgres database instead, for example in the server’s
.Renviron:
The tables are created on first use, and the record stays append-only: the database itself refuses to change or delete past entries. You need the RPostgres package installed.