Tsinghua's BaGuaLu-2 tames MoE expert-load spikes at runtime, accepted to SC26
jiqizhixin · x · 2026-11-12
BaGuaLu-2, from Tsinghua University and collaborators (accepted to SC26), treats MoE activation memory as a runtime control problem rather than a static pre-training configuration.
- Bottleneck: MoE routing can suddenly concentrate tokens on a few experts; static memory strategies either slow normal iterations or risk OOM on spikes
- Mechanism: after routing reveals load, a cost model picks operator-level activation mitigation (recomputation, offloading); a scheduler interleaves adjacent microbatches to overlap extra compute and transfers
- Payoff: average throughput reaches 1.14× Megatron-Kuai and 1.13× AdaP
More from Infra
- Polymarket prices 28% odds of a state data center moratorium; Torus raises $500M — Polymarket · 2026-11-12
- AWS-wrapper cloud startup shuts down just one month after seed round — thedealdirector · 2026-11-12
- Utah startup Torus raises $500M for shipping container-sized AI data centers deployable in 4 months — Polymarket · 2026-11-12
- Greylock reading group on AI accelerator design: more transistors no longer mean more performance — _shreya_s · 2026-11-12
- Dennard Scaling Is Dead: Automated Chip Design Could Shrink Hardware Cycles — _shreya_s · 2026-11-12
- Two smuggling scandals put Nvidia's China denials to the test — fortune · 2026-11-12