---

name: modeling-basics description: "Author and debug basic Pyro models, primitives, parameter state, validation, seeds, plates, and PyroModule parameters." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Modeling Basics for Pyro

Use this sub-skill when a task asks you to author or debug a basic Pyro program: stochastic functions, pyro.sample, pyro.param, observed data, pyro.plate, subsampling, parameter-store lifecycle, validation settings, RNG seeding, PyroModule, PyroParam, PyroSample, or pyro.render_model prerequisites.

This guidance targets the pyro-ppl 1.9.1 package family imported as pyro. It is self-contained for runtime use.

Route first

• For primitive usage, model/guide skeletons, observations, obs_mask, plates, validation, seeds, site names, and rendering prerequisites, read references/modeling-primitives.md.
• For pyro.param, ParamStoreDict, constraints, save/load state, pyro.module, PyroModule, PyroParam, PyroSample, and module_local_params, read references/modules-and-parameters.md.
• For warnings and failures such as duplicate sample names, observation warnings, parameter leakage, plate shape mismatches, missing Graphviz, or confusing validation toggles, read references/troubleshooting.md.

Reroute boundaries

• Deep distribution choice, event/batch/sample shape algebra, transforms, and support constraints: ../distributions-and-shapes/SKILL.md.
• SVI training loops, ELBO choice, autoguides, and optimizer recipes: ../svi-and-autoguides/SKILL.md.
• HMC/NUTS/MCMC, posterior predictive, and prediction APIs: ../mcmc-and-prediction/SKILL.md.
• Poutine handler composition, trace/replay/condition/block, enumeration, reparameterizers, and max_plate_nesting: ../effect-handlers-and-enumeration/SKILL.md.

Basic operating rules

1. Start repeated experiments, unit tests, and independent training runs with pyro.clear_param_store() unless you intentionally want to reuse existing parameters.
2. Keep validation on while developing: pyro.enable_validation(True). Use with pyro.validation_enabled(False): ... only around mature hot paths.
3. Seed reproducible snippets with pyro.set_rng_seed(seed) before generating synthetic data or initializing parameters.
4. Give every sample site a unique name within a single execution trace. Model and guide latent sites should share names; observed model sites usually have no guide site.
5. Use pyro.plate only for conditionally independent batch dimensions. If the error is mainly about batch_shape, event_shape, .to_event(), or log_prob shape, reroute to the shapes sub-skill.
6. Prefer PyroModule / PyroParam / PyroSample for module-style models. Use old pyro.module(name, nn_module, update_module_params=False) mainly for registering an ordinary torch.nn.Module with the global parameter store.

Entry points to recognize

• pyro.sample(name, fn, *args, obs=None, obs_mask=None, infer=None, **kwargs)
• pyro.param(name, init_tensor=None, constraint=constraints.real, event_dim=None)
• pyro.plate(name, size=None, subsample_size=None, subsample=None, dim=None, use_cuda=None, device=None)
• pyro.module(name, nn_module, update_module_params=False)
• PyroModule(name=""), PyroParam(init_value=None, constraint=constraints.real, event_dim=None), PyroSample(prior)
• pyro.settings.get(), pyro.settings.set(...), pyro.settings.context(...)
• pyro.enable_validation(...), pyro.validation_enabled(...), pyro.set_rng_seed(...)