It sounds like your use case is supported by Orchest today. Is model training dependent on GPU(s - multiple)?