python-numpy

python-numpy

The canonical forkd parent — Python 3.12 with numpy preinstalled. This is the image our front-page benchmark chart measures.

What you get

• python:3.12-slim base
• python3-numpy from apt (~30 MB, includes BLAS via OpenBLAS)
• forkd-init.sh + forkd-agent.py baked in as PID 1

Total rootfs: ~1.5 GB, memory image after warm-up: ~512 MiB.

Use it

sudo bash recipes/python-numpy/build.sh
sudo bash scripts/host-tap.sh                # one-time host tap setup
sudo forkd snapshot --tag pyagent \
    --kernel ./vmlinux-6.1.141 \
    --rootfs recipes/python-numpy/parent.ext4 \
    --tap forkd-tap0

# Fork 100 sandboxes; each can `eval()` numpy expressions instantly.
sudo bash scripts/netns-setup.sh 100
sudo -E forkd fork --tag pyagent -n 100 --per-child-netns
sudo forkd eval --child forkd-child-1 -- "numpy.zeros(5).tolist()"
# [0.0, 0.0, 0.0, 0.0, 0.0]

When to pick this

• You want to reproduce the benchmark numbers.
• Your workload uses NumPy / SciPy but doesn't need a full ML stack.
• You want the smallest possible Python image with one heavy import already warmed.

For heavier ML (PyTorch, TensorFlow), build a custom rootfs from a PyTorch base image — see coding-agent/ for the pattern.