Use a model from Rust

This page covers saving a model built in Python and loading it in a Rust program, so inference runs without a Python interpreter. You need it if you build models in Python but deploy into a Rust service; if you work only in Python, skip it.

Python and Rust share the rustmc.graph-model version 1 format. The definition, expression compiler, schema validation, and loader live in rustmc_core::model. Python models saved by 0.12 remain readable.

Save a compiled Python model with Path("model.json").write_text(compiled.to_json()). In Rust, use GraphModel::from_json, bind DataInputs, then call sample or log_density. ModelFit::predict accepts future inputs without response placeholders. Its output is indexed by response, chain, draw, and observation. Displayed samples preserve transformed and noncentered parameter identities.

The runnable load_model example accepts four JSON files: the model artifact, training data, an unconstrained parameter vector, and future predictors. Data files use the same keyed vectors and row-major matrices as Python dictionaries.

cargo run -p rustmc_core --release --example load_model -- \
  model.json data.json position.json future.json

The example prints a log density, gradient, posterior means, and conditional-mean predictions. It uses fixed demonstration sampling controls; applications should supply their own SamplerConfig and inspect diagnostics.

Compiled artifacts contain structure, not training data. Fitted artifacts (rustmc.graph-fit) contain training data and draws, and load on either side: a fit saved with fit.to_json() in Python opens in Rust with ModelFit::from_json, which validates every stored draw against the model, and ModelFit::to_json writes a file Python's FitResult.from_json reads. Both write keys in a fixed order, so the same fit saves to the same bytes. ModelFit also provides log_likelihood, deterministics and posterior_predictive. GraphModel::sample_batch fits many datasets against one model. Neither format checkpoints sampler state. The older data-owning CompiledModelArtifact format and its rustmc_core::compiled_model module have been removed, so rustmc.graph-model is the only compiled-model artifact.

The Rust API is alpha. The loader rejects unknown versions, and every artifact type it deserializes — ModelArtifact, ModelSpec, PriorSpec, LikelihoodSpec, MuExpr, DataSchema, DataSlot, and SlotKind — denies unknown fields. A field this version does not recognise is an error rather than a silent drop, so an artifact written by a newer version fails to load instead of loading as a different model.