atomli-tako-ci

The Tako Code Interpreter runtime, packaged for PyPI, with a minimal slice of the calculation WASM cores and model checkpoints bundled inside the wheel.

It is the pip sibling of the npm package @atomli/tako-ci. Sandboxes such as ChatGPT Code Interpreter cannot fetch arbitrary URLs — the only inbound channel is the package proxy — so Tako delivers its runtime and the assets it needs as an installable package instead of an over-the-wire download.

Install

pip install atomli-tako-ci

Run a job

python3 -m tako_ci job.json --manifest runtime-manifest.json --output result.tako.zip

job.json and runtime-manifest.json are produced by Tako's prepare_tako_script tool. The runner loads the WASM core (and, for MLIP/Skala jobs, the model checkpoint), executes the calculation on a Node-compatible host (node or bun), and writes a result.tako.zip bundle.

What ships in the wheel

The runner (tako_ci/tako_ci.py) resolves every asset named in the runtime manifest by first looking for a bundled copy under tako_ci/assets/, verifying its SHA-256, and only falling back to a signed network download when the asset is not present locally.

To stay within a git- and PyPI-friendly size (~35 MB), all three WASM cores plus the small checkpoints are bundled, while the large model weights are left out:

Only the large model weights are left out — the nequip-l checkpoint (72 MB) and the equiformer / equiformer-gradient checkpoints (486–560 MB). Jobs that need them fetch them over the network in hosts that allow it; in fetch-restricted sandboxes those jobs run in the full Tako application instead.

Bundled digests are pinned in tako_ci/manifest.json.