Machine Heart
Oct 27, 2026 · Artificial Intelligence
BaGuaLu-2: Adaptive Activation Memory Management for MoE Training at Scale
Tsinghua's BaGuaLu-2, accepted at SC26, introduces runtime adaptive activation memory provisioning that dynamically selects recomputation and offloading per operator to handle MoE memory spikes on 12,288-GPU clusters, boosting throughput 1.14× and achieving 32.9% MFU without manual configuration.
BaGuaLu-2Huawei DCUMoE training
0 likes · 9 min read
