The Host Is Not Idle: Computational SSDs for Host-Free Checkpointing in Multimodal LLM Training
Large language models (LLMs) are increasingly extending beyond plain text to jointly reason over images and video. Among such multimodal models, video LLMs that take video as input require the host CPU to decode and preprocess each upcoming batch of video, making it busy throughout training. We show that this property...
Seungkwan Kang, Hyungseok Ko, Hyun Dong Lee et al.
· Proceedings of the 18th ACM... · 0 citations