Train in cluster, infer anywhere

How Krea shares GPUs between research and production with a custom solution that lets inference scale anywhere when training claims every GPU.