add-model-bundle

To add a new diffusion model package to UniRL, you scaffold the expected package directory structure, define typed dataclass configs, and implement the bundle loader and stage classes for framework discovery and instantiation.

A UniRL model bundle is a structured package integration that defines pipeline wiring, typed conditions, and rollout integration for new diffusion or autoregressive models, needed when adding new model support to the framework without reverse-engineering patterns.

UniRL model bundle support requires implementing FSDP hints and LoRA targets within the conditioning and runtime integration, alongside meta-init behavior and rollout plumbing to ensure proper framework functionality.

To validate a new autoregressive model integration in UniRL, implement targeted condition, step, stage, and pipeline tests to ensure the new model support remains reliable and safely integrated into the framework.

The best way to update UniRL pipeline configs is using plain dataclass configs with hydra-config support, defining pipeline config fields, export points, and recipe updates across the UniRL model families for safe integration.