@misc{indiciae9bd143ff6f24,
  title = {A Training-Memory Regression in MLA Sequence Parallelism: Why Megatron-Core Forbids Absorption, and LAGA -- a Communication-Efficient Fix},
  author = {Changzheng Ma},
  year = {2026},
  url = {https://arxiv.org/abs/2607.17644},
  note = {Source identifier: 2607.17644}
}