Tags
1 page
Distributed Training
Megatron Tensor Parallelism: Rebuilding a SwiGLU Feed-Forward Block Across Two Ranks