Skip to content

Latest commit

 

History

History
127 lines (93 loc) · 4.17 KB

README.md

File metadata and controls

127 lines (93 loc) · 4.17 KB

ACiD Build Status Latest

This is an implementation of an asynchronous multiprocessing optimization algorithm with a continuous local momentum called A²CiD² in the Julia programming language as introduced in [1].

Note

There is also an official demo by the original authors of [1] in Python: AdelNabli/ACiD

Usage

Install mpiexecjl

You can install mpiexecjl with MPI.install_mpiexecjl(). The default destination directory is joinpath(DEPOT_PATH[1], "bin"), which usually translates to ~/.julia/bin, but check the value on your system. You can also tell MPI.install_mpiexecjl to install to a different directory.

$ julia
julia> using MPI
julia> MPI.install_mpiexecjl()

To quickly call this wrapper we recommend you to add the destination directory to your PATH environment variable.

References

[1] A. Nabli, E. Belilovsky, and E. Oyallon, “A²CiD²: Accelerating Asynchronous Communication in Decentralized Deep Learning,” in Thirty-seventh Conference on Neural Information Processing Systems, 2023. [Online]. Available: arXiv DOI:10.48550/ARXIV.2306.08289