HiLoRe: What to Store, Compress, or Recompute for Efficient GRPO Training
Group-relative policy optimization (GRPO) makes learner-side activations a major memory-computation bottleneck: gradient checkpointing reduces activation memory through recomputation, but fixed schedules can leave roughly 18 GB unused on a 48-GB GPU despite substantial recomputation overhead. Existing activation-manage...
Xin-Rui Chen, Mengyang Li, Ou Wu et al.
· 0 citations