DeepSeek — Field Evidence

Communication-Aware Placement and Pruning for Efficient Mixture-of-Experts Inference — As MoE models scale to hundreds of experts, placement and pruning decisions increasingly dictate communication volume, affecting the performance of distributed inference across GPUs and nodes.