Drop-Upcycling: Training Sparse Mixture of Experts with Partial Re-initialization.
Taishi Nakamura, Takuya Akiba, Kazuki Fujii, Yusuke Oda, Rio Yokota, Jun Suzuki
Browse the full ICLR paper archive.
Taishi Nakamura, Takuya Akiba, Kazuki Fujii, Yusuke Oda, Rio Yokota, Jun Suzuki
Browse the full ICLR paper archive.