method · first seen 9 Oct 2026
activation swapping
Also: activation offloading
Moving saved activations from GPU memory to host memory during training and fetching them back when needed.
Mentions per week
PapersCodeDiscussion
Earliest sources
- SlackSwap: Bandwidth-Aware Activation Swapping and Recomputation for Memory-Efficient DNN TrainingPapers · 9 Oct 2026 · Hui Feng, Xun Zhou, Yuxin Wang
- OOM: Offload-Optimized Full-Parameter MoE Training via Memory-Execution Co-DesignPapers · 8 Nov 2026 · Ruijia Yang, Shiyuan Lin, Chengxi Liao, Zeyi Wen
Latest sources
- OOM: Offload-Optimized Full-Parameter MoE Training via Memory-Execution Co-DesignPapers · 8 Nov 2026 · Ruijia Yang, Shiyuan Lin, Chengxi Liao, Zeyi Wen
- SlackSwap: Bandwidth-Aware Activation Swapping and Recomputation for Memory-Efficient DNN TrainingPapers · 9 Oct 2026 · Hui Feng, Xun Zhou, Yuxin Wang