Fails to load (OOM) models that I can load fine in LMStudio - probably related to worse (or potentially nonexistent?) automatic handling of how many GPU layers to offload