Abstract
High-resolution video diffusion models built on Diffusion Transformers (DiTs) deliver strong fidelity but quickly exhaust the memory budget of single-workstation setups. A DiT with over 100 billion parameters can require more than a terabyte of persistent state, while naive spatiotemporal self-attention grows quadratically with sequence length. These two constraints—parameter memory and activation memory—prevent researchers from adapting large generative models without access to large GPU clusters.
Revisiting this problem from a systems perspective, we introduce MegaSlide-DiT, a prototype that demonstrates how a pre-trained 105-billion parameter DiT can be adapted on a single H200 GPU equipped with 1.5 TB of host RAM. Our key insight is that the GPU does not need to own the entire model state: all persistent weights, master weights, and optimizer moments remain in host memory, while only transient shards are streamed to the GPU on demand. At the same time, we replace quadratic global attention with 3D Deformable Slide Attention (3D-DSA), a motion-adaptive local attention operator that reduces both memory and computational complexity to linear in sequence length.
We provide detailed memory accounting, execution traces, and evaluation results to substantiate our design. MegaSlide-DiT does not claim to train a 105-billion parameter model from scratch on a single GPU, nor does it magically solve bandwidth limits; instead, it offers a pragmatic path for full-parameter adaptation of massive video diffusion models on high-end workstations.
Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
Cite as: arXiv:2607.22696 [cs.CV]
Submitted on: 17 July 2026
Authors: Jiacheng Liu, Jason Liu
DOI: https://doi.org/10.48550/arXiv.2607.22696
via ArXiv CV