1 paper · 1 filter
Ping Chen, Wenjie Zhang, Shuibing He +10
Large model training often uses recomputation to alleviate memory pressure and pipelines to exploit the parallelism of data, tensors, and devices. However, existing recomputation a…