AI Paper Atlas

Research map / Optimization / Distributed and compressed

DeToNATION: Decoupled Torch Network-Aware Training on Interlinked Online Nodes

Mogens Henrik From, Jacob Nielsen, Lukas Galke, Peter Schneider-Kamp

AAAI 2026 · Optimization › Distributed and compressed · approach: Gradient quantization and sparsification · 0 citations (OpenAlex)

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace DeMo DCT replication with Random replication (same compression)FlexDeMo-SGD · T5 Opus Books Fr-Envalidation lossImproved
replace Random with DeMo replicationFlexDeMo-SGD · ViT-B CIFAR-100validation lossImproved
replace Random with DeMo replicationFlexDeMo-SGD · OLMo2 1B Dolma v1.6train lossImproved
use Striding replication instead of DeMoFlexDeMo-SGD · OLMo2 1B Dolma v1.6train lossHurt
use DiLoCo replication instead of DeMoFlexDeMo-SGD · T5 Opus Books Fr-Envalidation lossHurt
share only the sign of the gradient components (ablation)DeMo replication · T5, ViT, OLMo hyperparameter study (Appendix)loss convergenceImproved
use fp32 instead of fp16 transfer dtype (ablation)DeMo replication · Hyperparameter study (Appendix)loss convergenceImproved
increase chunk size / TopK (ablation)DeMo replication · Hyperparameter study (Appendix)loss convergenceMixed
use FlexDeMo instead of Hybrid-FSDP with full-gradient AdamWHybrid-FSDP · OLMo2 1B Dolma v1.6, 2 nodeswall-clock time: about 2.6 times fasterImproved
use FlexDeMo instead of Hybrid-FSDP with full-gradient AdamWHybrid-FSDP · OLMo2 1B Dolma v1.6, 2 nodestrain lossImproved
replace Random with DeMo replication at 64 nodesFlexDeMo-SGD · OLMo2 1B, 64 nodeswall-clock timeHurt
use Random instead of DeMo replication at 64 nodesFlexDeMo-SGD · OLMo2 1B, 64 nodestrain loss: 64% faster than conventional setupHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy