DeepSeek Infrastructure Profiling Data Released

DeepSeek is publicly sharing profiling data from its training and inference framework to help the community understand its communication-computation overlap strategies and low-level implementation details. The data, captured using the PyTorch Profiler, can be visualized directly in Chrome or Edge browsers. The analysis simulates a perfectly balanced MoE routing strategy and covers training, prefilling, and decoding phases. Different configurations (e.g., EP64/TP1, EP32/TP1, EP128/TP1) and micro-batching strategies are optimized for computation and communication overlap to improve efficiency.