论文精选10:03ParaTTT算法将测试时训练层的非线性记忆并行化TTT层一直快不起来，这篇给出并行化算法，比TTT-MLP快32倍，训练340M模型提速4.9倍，还追平Mamba-2。#ParaTTT#TTT#Mamba-2#Gated DeltaNetaarXiv cs.LG@Noel Loo 等 6 人原文稍后读已读值得跟进有用关注 ParaTTT