2 papers
cs.LG2026
HiLoRe: What to Store, Compress, or Recompute for Efficient GRPO Training
Xinrui Chen, Mengyang Li, Ou Wu +1
Group-relative policy optimization (GRPO) makes learner-side activations a major memory-computation bottleneck: gradient checkpointing reduces activation memory through recomputati…
cs.LG2022
Class-Level Logit Perturbation
Mengyang Li, Fengguang Su, Ou Wu +1
Features, logits, and labels are the three primary data when a sample passes through a deep neural network. Feature perturbation and label perturbation receive increasing attention…