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](https://runtimewire.com/author/runtimewire-staff)
· Published
· Updated
Hugging Face has fixed a training-correctness bug in TRL, its open-source model-training library, that let asynchronous reinforcement-learning rollouts reuse vLLM prefix-cache blocks computed under older model weights. The patch, pull request #8300, adds 'cache_salt=f"v{model_version}"' to affected rollout requests so each policy version gets its own prefix-cache namespace. The change landed in TRL v1.18.0, whose release notes describe the stale prefix KV as lasting for the whole run before the fix.…
source & further reading
runtimewire.com — original article
Cloudflare announces open-source Kitesurf browser engine for AI agents
Six Chinese AI models carry Blackwell fingerprints in their published weights, researcher says
After 3.5 years, llama.cpp finally calls itself 1.0