Mert Toslali

### Your current environment I am utilizing vLLM sleep in HF - TRL to efficiently manage GPU memory between training and generation. See my draft [PR](https://github.com/toslali-ibm/trl/pull/4/files). The training is completed...