Verl v0.9.1 introduces breaking changes to default behavior

The release of verl-project/verl v0.9.1, as detailed in the project's GitHub release notes, includes several changes that alter default behavior and may impact existing systems. Notable updates include the integration of a unified V1 trainer, modifications to the dynamic micro-batch packing mechanism, and the enforcement of a strict max_token_len limit. Additionally, the dependency management system has transitioned to uv, replacing previous methods with pyproject.toml and a committed uv.lock file. Multimodal Continuous Token functionality has been revamped, altering the ContinuousTokenBuilder interface and introducing new composition methods. Other breaking changes include the removal of the mindspeedllm backend engine, default adjustments for algorithm.filter_groups.metric, and the introduction of new configurations for V1 async trainers.

Operators should carefully review their configurations before upgrading, particularly if they rely on the removed mindspeedllm backend or the deprecated fully_async_policy and one_step_off_policy modules. The shift to uv for dependency management may require adjustments to installation workflows, especially for environments not yet aligned with uv standards. Systems utilizing multimodal processing or dynamic micro-batching should validate their token limits and builder compatibility to avoid unexpected behavior. These changes reflect a broader trend in projects to consolidate trainer functionality while introducing stricter enforcement of resource and configuration boundaries.