DeepSeek Open Source Optimized Parallelism Strategies, 3 repos

The GitHub repository deepseek-ai/profile-data publicly shares profiling data for training and inference frameworks, showcasing strategies for computational and communication overlap, mainly focusing on MoE layers within the DeepSeek infrastructure.

• Public profiling data for training frameworks.
• Uses PyTorch Profiler for data capture.
• Visualized via Chrome or Edge tracing.
• Affiliated with DualPipe and DeepEP.
• Balanced MoE routing strategy simulated.