Building custom training pipelines on distributed infrastructure forces you to bend your research to fit the framework's constraints. Most distributed training frameworks treat custom loss functions, non-standard architectures, and experimental optimizers as afterthoughts rather than first-class citizens. The framework dictates your research direction instead of adapting to your experimental needs. This is a fundamental design flaw in current distributed training tools, they optimize for common cases and leave edge cases (which are often the most innovative research directions) poorly supported or completely broken.