Learned Subspace Compression for Communication-Efficient Pipeline Parallelism

In three linesMAPL (Manifold Aware Projection Learning) compresses inter-stage activations in pipeline parallelism using learned orthogonal projections constrained to Stiefel manifold. Tested on LLaMA 150M–1B, it achieves high compression with negligible performance loss, outperforming Subspace Networks.

Summary generated by Claude — human-verified