Describe how communication overhead such as All-Reduce impacts distributed training scaling efficiency at Mistral AI, and how you would mitigate it.
