We identified a KV cache gradient leak inherited from Self Forcing: no_grad() refreshes cache values but can retain the old autograd graph.
A two-line detach fix cut DMD training memory by 22%.
Using this training pattern? Check your cache 📷
linydthu.github.io/blog/2026…
9
Donglin Yang retweeted
Ever wondered how card shuffling can help solve combinatorial problems like sorting, jigsaw puzzles, or TSPs?
Our ICLR 2025 (Oral) paper SymmetricDiffusers introduces a novel discrete diffusion model on finite symmetric groups to solve them in a unified and principled way!
#ECCV2022
If you are interested in how to improve the behavioral cloning with observation histories, please check out our latest work on ECCV2022!!
Behavioral cloning+observation histories > behavioral cloning+the most recent observation? Surprisingly, it sometimes won’t. This counter-intuitive phenomenon is referred to as copycat problem. Our ECCV2022 work proposes a novel imitation learning network to alleviate this issue.




