Categories
1 page
Deep-Learning Mechanisms
Megatron Tensor Parallelism: Rebuilding a SwiGLU Feed-Forward Block Across Two Ranks