Detailed profiling data from a training and inference framework is shared, highlighting communication-computation overlap strategies with PyTorch Profiler visualizations. The framework implements DualPipe with MoE layers across different configurations, including EP64/TP1 for training and EP32/TP1 for prefilling, demonstrating balanced routing and micro-batch optimization techniques.
DualPipe is a bidirectional pipeline parallelism algorithm that optimizes computation-communication overlap in neural networks by achieving full overlap of forward and backward phases. The solution, presented in the DeepSeek-V3 Technical Report, reduces pipeline bubbles and requires implementation of custom overlapped forward-backward methods for specific modules.
FlashMLA is a high-performance MLA decoding kernel optimized for Hopper GPUs, achieving up to 3000 GB/s in memory-bound configurations and 580 TFLOPS in computation-bound scenarios. The implementation supports BF16 and paged kvcache, requiring CUDA 12.3+ and PyTorch 2.0+.