---

name: svi-autoguides description: "Fit and debug NumPyro stochastic variational inference workflows with ELBOs, manual guides, autoguides, optimizers, and enumeration choices." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

svi-autoguides

Use this sub-skill when the task is to fit a NumPyro model with stochastic variational inference (SVI), write or repair a guide, choose an ELBO, use automatic guides, inspect learned parameters and losses, handle discrete latent variables in SVI, or turn SVI results into posterior/predictive samples.

Quick workflow

1. Confirm model and guide signatures. Model/guide primitive and shape issues route through ../modeling-primitives/ and distribution support issues through ../distributions-transforms/.
2. Choose a guide path. Use a manual guide for custom variational families or constraints; use references/autoguide-reference.md for AutoNormal, AutoDelta, normal-family autoguides, Laplace, flow guides, DAIS, and guide lists.
3. Choose an ELBO. Use references/elbo-and-enumeration.md for Trace_ELBO, TraceMeanField_ELBO, TraceGraph_ELBO, TraceEnum_ELBO, RenyiELBO, and discrete enumeration.
4. Run SVI. Use references/svi-workflows.md for SVI.init, update, run, get_params, evaluate, optimizer selection, and posterior predictive use.
5. Validate losses and params. Check finite losses, constrained parameter values, expected keys, and posterior predictive shapes.
6. Run the bundled smoke. scripts/svi_smoke.py is a small synthetic SVI check adapted from NumPyro's minimal example.
7. Triage failures. Use references/troubleshooting.md for NaN losses, invalid guide values, wrong ELBO for discrete sites, missing optax/funsor, and local-latent guide limitations.

Prefer this sub-skill for

• SVI(model, guide, optim, loss) setup and update loops.
• Trace_ELBO, TraceMeanField_ELBO, TraceGraph_ELBO, TraceEnum_ELBO, or RenyiELBO decisions.
• numpyro.param constraints in guides.
• AutoNormal, AutoDiagonalNormal, AutoMultivariateNormal, AutoLowRankMultivariateNormal, AutoDelta, AutoGuideList, AutoLaplaceApproximation, AutoIAFNormal, AutoBNAFNormal, AutoDAIS, AutoSemiDAIS, and AutoSurrogateLikelihoodDAIS.
• Discrete enumeration with config_enumerate, infer={"enumerate": "parallel"}, and TraceEnum_ELBO.
• Optimizer selection with numpyro.optim or optional Optax optimizers.

Route elsewhere when

• The task is MCMC/NUTS/HMC, chain diagnostics, divergences, or posterior samples from MCMC: use ../mcmc-diagnostics/.
• The task is a core model/handler/plate trace before fitting: use ../modeling-primitives/.
• The task is distribution constructor/support/transform shape semantics: use ../distributions-transforms/.
• The task is SteinVI/SVGD/ASVGD, Funsor internals, neural module wrappers, nested sampling, HSGP, or TFP: use ../advanced-contrib/.

Bundled materials

• SVI workflows covers manual guides, loops, params, losses, and predictive use.
• Autoguide reference maps guide families to task signals and caveats.
• ELBO and enumeration explains loss selection and discrete latent handling.
• Troubleshooting gives symptom-to-fix guidance.
• SVI smoke script validates a tiny synthetic variational fit.