Hugging Face fixes TRL bug that let RL rollouts reuse KV cache from older weights
TRL v1.18.0 adds a model-version cache salt so AsyncGRPO and AsyncDistillation student rollouts stop pairing new weights with first-version prefix KV.
By RuntimeWire Staff · Published · Updated

TRL v1.18.0 adds a model-version cache salt so AsyncGRPO and AsyncDistillation student rollouts stop pairing new weights with first-version prefix KV.