---

name: core-models description: "Select and use scvi-tools built-in core model families, including setup, training, and common accessors for SCVI, SCANVI, TOTALVI, PEAKVI, MULTIVI, AUTOZI, LinearSCVI, CondSCVI, DestVI, AmortizedLDA, and mlxSCVI." disable-model-invocation: true metadata: disco-role: operating license: BSD 3-Clause

Core Models

Use this sub-skill when choosing among scvi-tools built-in model families or writing common scvi.model workflows. It covers model selection, modality requirements, setup/train patterns, core constructor and setup_anndata signatures, and failures around model registration or modality mismatch.

Fast Routing

• Need to choose a model from data modalities or labels: read references/model-selection.md.
• Need exact core setup_anndata, constructor, or train parameters: read references/api-reference.md.
• Need to debug missing setup, wrong keys, protein/ATAC/RNA modality shape issues, or optional MLX/Pyro dependencies: read references/troubleshooting.md.
• Need to inspect the installed package API rather than trust static notes: run python scripts/inspect_model_api.py --model SCVI --model TOTALVI from this sub-skill directory; repeat --model for each model.

Boundaries

• Use the sibling data-setup skill for loading datasets, AnnData/MuData preparation, count layers, registry internals, and setup schema design.
• Use the training-and-inference skill for accelerator selection, callbacks, logging, checkpointing, custom Trainer options, and long-running inference operations.
• Use downstream-analysis for differential expression/accessibility, latent visualization, normalized expression/accessibility interpretation, clustering, and result analysis.
• Use multimodal-and-spatial or annotation-and-query for external specialized models and deep spatial/query transfer workflows beyond the built-in core family overview.

Minimal Core Pattern

import scvi

scvi.model.SCVI.setup_anndata(adata, layer="counts", batch_key="batch")
model = scvi.model.SCVI(adata, n_latent=10)
model.train(max_epochs=100)
latent = model.get_latent_representation()

Always call the matching class's setup_anndata on the object before constructing that class, and pass modality-specific keys such as protein_expression_obsm_key only when the model signature supports them.