{"id":30198984,"url":"https://github.com/shigangli/eager-sgd","last_synced_at":"2025-08-13T07:39:08.288Z","repository":{"id":53279024,"uuid":"225068615","full_name":"Shigangli/eager-SGD","owner":"Shigangli","description":"Eager-SGD is a decentralized asynchronous SGD. It utilizes novel partial collectives operations to accumulate the gradients across all the processes.","archived":false,"fork":false,"pushed_at":"2021-11-18T16:14:30.000Z","size":1376,"stargazers_count":7,"open_issues_count":1,"forks_count":0,"subscribers_count":3,"default_branch":"master","last_synced_at":"2023-10-20T23:06:23.717Z","etag":null,"topics":["distributed-deep-learning","gradient-averaging","partial-allreduce"],"latest_commit_sha":null,"homepage":"","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"apache-2.0","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/Shigangli.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null}},"created_at":"2019-11-30T20:53:07.000Z","updated_at":"2023-10-20T23:06:23.718Z","dependencies_parsed_at":"2022-08-28T15:50:39.806Z","dependency_job_id":null,"html_url":"https://github.com/Shigangli/eager-SGD","commit_stats":null,"previous_names":[],"tags_count":0,"template":null,"template_full_name":null,"purl":"pkg:github/Shigangli/eager-SGD","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Shigangli%2Feager-SGD","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Shigangli%2Feager-SGD/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Shigangli%2Feager-SGD/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Shigangli%2Feager-SGD/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/Shigangli","download_url":"https://codeload.github.com/Shigangli/eager-SGD/tar.gz/refs/heads/master","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Shigangli%2Feager-SGD/sbom","host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":270204953,"owners_count":24544683,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","status":"online","status_checked_at":"2025-08-13T02:00:09.904Z","response_time":66,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["distributed-deep-learning","gradient-averaging","partial-allreduce"],"created_at":"2025-08-13T07:39:04.754Z","updated_at":"2025-08-13T07:39:08.239Z","avatar_url":"https://github.com/Shigangli.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"Eager-SGD\n---------\n\n**Eager-SGD** is a **decentralized asynchronous SGD** for distributed deep learning training based on **gradient averaging**. It utilizes novel partial collectives operations (partial allreduce) to accumulate the gradients across all the processes. Different from the traditional collectives operations (such as MPI, NCCL), a partial collective is an asynchronous operation where a subset of the processes can trigger and contribute the latest data to the collective operation. \n\nEager-SGD may bring staleness to the gradients. Thanks to our sophisticated implementation of solo-allreduce and majority-allreduce, the **staleness is bounded** and therefore eager-SGD is stale-synchronous. Due to the asynchrony feature of eager-SGD, it can better handle the deep learning training with load imbalance. To the best of our knowledge, this is the first work that implements asynchronous and stale-synchronous decentralized SGD where the messages propagate to all nodes in one step.\n\nDemo\n---------\nA script to run eager-SGD on ResNet-50/ImageNet with SLURM job scheduler can be found [here](https://github.com/Shigangli/eager-SGD/blob/master/test-models/tf-models-r1.11/official/resnet/test_scripts_imagenet/daint_eagersgd_imagenet.sh).\nGenerally, to evaluate other neural network models with the [customized optimizers](https://github.com/Shigangli/eager-SGD/blob/master/test-models/tf-models-r1.11/official/utils/) (e.g., gradient averaging using solo/majority-allreduce), one can simply wrap the default optimizer using the customized optimizers. See the example for ResNet-50 [here](https://github.com/Shigangli/eager-SGD/blob/master/test-models/tf-models-r1.11/official/resnet/resnet_run_loop_solo_imagenet_300.py#L384).\n\n\n\nPublication\n-----------\nThe work of eager-SGD is pulished in PPoPP'20, **Best Paper Finalist**. See the [paper](https://shigangli.github.io/files/ppopp20-eager-SGD-paper.pdf) for details. If you use eager-SGD, cite us:\n```bibtex\n@inproceedings{li2020taming,\n  title={Taming unbalanced training workloads in deep learning with partial collective operations},\n  author={Li, Shigang and Ben-Nun, Tal and Girolamo, Salvatore Di and Alistarh, Dan and Hoefler, Torsten},\n  booktitle={Proceedings of the 25th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming},\n  pages={45--61},\n  year={2020}\n}\n```\n\nLicense\n-------\nSee [LICENSE](LICENSE).\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fshigangli%2Feager-sgd","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fshigangli%2Feager-sgd","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fshigangli%2Feager-sgd/lists"}