Reproduce “THC: Accelerating Distributed Deep Learning Using Tensor Homomorphic Compression” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Minghao Li, Ran Ben Basat, Shay Vargaftik, ChonLam Lao, Kevin Xu, Michael Mitzenmacher, Minlan Yu

THC makes compressed gradient representations directly aggregable, allowing a parameter server to combine and return them without decompressing and recompressing each round. Workers use randomized Hadamard transforms, quantization, and error feedback; programmable switches can aggregate the values. On a four-worker testbed, THC with a Tofino switch reached GPT-2’s 81% target accuracy 1.47 times faster than Horovod-RDMA for evaluated communication-sensitive workloads.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.