znyx-inference 1.0.0

Znyx inference sidecar: an optional HTTP service that serves ML models for model-backed guardrail detection. Boots dependency-free on a stub runner; add the [onnx] extra to serve real weights on CPU.

znyx-inference

The optional Znyx inference sidecar: an HTTP service that serves ML models for model-backed guardrail detection. The znyx-runtime calls it over HTTP to escalate from rules to ML. Without it, the runtime still works fully on its deterministic rules path.

# boots dependency-free on a stub runner (no ML stack needed):
pip install znyx-inference

# to serve real model weights on CPU — lean, torch-free (onnxruntime + tokenizers):
pip install "znyx-inference[onnx]"

# offline only: export + quantize a checkpoint to a pinned ONNX artifact (heavy: torch + optimum):
pip install "znyx-inference[export]"

Weights are never bundled

This package ships code only, no model weights. You fetch the models you want explicitly, sha256-pin them, and mount them. Nothing is downloaded implicitly on the serving path. See MODELS.md for the curated model list, licenses, and the fetch-and-pin workflow, including which models carry special licenses.

How it fits

your app -> znyx-runtime (rules) --HTTP--> znyx-inference (ML) -> local, pinned weights

The sidecar loads only local, sha256-verified weights, fails closed, and never phones home. If a model is not present it stays on the stub runner and the runtime falls back to rules per the policy's fallback mode.

See the repository README for the full architecture.

License

Apache-2.0. Model weights you fetch are governed by their own licenses (see MODELS.md).