We have to disable GPU train tests for a few models, because their original batch size doesn't fit into our current CI infra. List of models that disabled train because of insufficient GPU memory: - densenet121 - demucs - hf_T5 - squeezenet1_1 - tacotron2 - timm_nfnet