CLiS: Plug-and-Play Attention Acceleration for Video Diffusion on TPUs via Certified Stabilization
cs.LG
Submitted: 2026-10-30
Updated: 2026-10-30
Code: https://github.com/google/torchax
License: http://creativecommons.org/licenses/by/4.0/
The gist: High-resolution video diffusion relies on attention over long token sequences, making attention efficiency critical to generation latency.
Terminology
Abstract
High-resolution video diffusion relies on attention over long token sequences, making attention efficiency critical to generation latency. Our analysis identifies repeated maximum reductions and accumulator rescaling in online softmax as a bottleneck in long-context TPU attention. We propose CLiS (Certified Loop-invariant Stabilization), a training-free, plug-and-play optimization for video diffusion inference. CLiS fixes per-row stabilizers before the KV loop and certifies numerical safety per head, falling back to online softmax when certification fails. It integrates into existing fused-attention kernels without changing model weights or architecture. On Wan 2.2 and LTX-2 across TPU v6e and TPU7x, CLiS reduces diffusion transformer (DiT) step latency by up to 19.6% relative to tuned online softmax without measurable quality loss. Controlled TPU adaptations of prior GPU stabilization methods show performance or numerical-safety limitations. Preliminary causal LLM experiments also show up to 13.5% higher training throughput in continued pretraining and 7.7% in supervised fine-tuning.
Sources
- FLASH-D: FlashAttention with Hidden Softmax Division
- DeepSeek-V3 Technical Report
- Nemotron-CLIMB: CLustering-based Iterative Data Mixture Bootstrapping for Language Model Pre-training
- Gemma 3 Technical Report
- The Llama 3 Herd of Models
- LTX-Video: Realtime Video Latent Diffusion
- LTX-2: Efficient Joint Audio-Visual Foundation Model
- Ragged Paged Attention: A High-Performance and Flexible LLM Inference Kernel for TPU
- A Study of BFLOAT16 for Deep Learning Training
- HunyuanVideo: A Systematic Framework For Large Video Generative Models
- HyperVAttention: Efficient Sparse Attention with Spatio-Temporal Clustering for Video Diffusion
- Online normalizer calculation for softmax
- VFA: Relieving Vector Operations in Flash Attention with Global Maximum Pre-computation
- Wan: Open and Advanced Large-Scale Video Generative Models
- Qwen3 Technical Report
- UniPC: A Unified Predictor-Corrector Framework for Fast Sampling of Diffusion Models
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis