v0.12.1
v0.12.1 (2026-03-26)
This release is published under the Apache-2.0 License.
Bug Fixes
A customer reported that their training codebase uses protobuf 5 and hit a version conflict with cwsandbox requiring >=6.33.5. An audit of ~25 ML training frameworks shows protobuf v5 is the common ground for PyTorch-based stacks: most don't constrain protobuf at all, and the few that do (Composer <5.30, vLLM >=5.29.6) target v5.
Frameworks audited include DeepSpeed, Megatron-LM, TRL, Accelerate, Transformers, torchtitan, verl, NeMo, Composer, LLM Foundry, Axolotl, LLaMA-Factory, Unsloth, vLLM, SGLang, Ray, PyTorch Lightning, PEFT, OpenRLHF, W&B, TensorBoard, and others. Only TensorFlow HEAD requires protobuf v6+, but TF environments are separate from PyTorch training.
Use buf.build plugin version 26.1.0 (protobuf 5.26.1) to generate proto stubs that predate the ValidateProtobufRuntimeVersion check introduced in 5.27.0. This removes the runtime version pinning from generated stubs, allowing protobuf >=5,<7 as the dependency range. Both protobuf 5.x and 6.x pass the full unit test suite. A CI matrix job verifies compatibility with both major versions.
Detailed Changes: v0.12.0...v0.12.1