Summary
verl 0.9.1 adds a rollout-router protocol and factory, extends delta-sharded weight synchronization to TorchTitan, and introduces GPU-lending async training. The release also moves dependency management to pyproject.toml and a committed uv.lock, making uv the supported installation path.
Why it matters
Research and engineering teams operating large agent-training jobs gain new extension points and lower synchronization overhead between training and generation workloads.
Related coverage
- OpenAI Node SDK 7.20.0 adds vault credentials and safety lifecycle APIs
- OpenAI Go SDK 3.64.0 adds agent safety and session-environment controls
- Qwen Code 0.24.2 adds a Linux sandbox foundation and durable workflow controls