Cerebras uses μP, DeepSeek does not. How would you choo...

The choice depends on your research goals. μP requires a very clean implementation, including per-layer learning rates and specific attention scaling, but it pays off significantly if you are performing multiple scaling experiments. If your project involves training only one model at a fixed target scale, the overhead of implementing μP may not be necessary, and direct hyperparameter fitting is a simpler, equally effective approach.