arXiv:2609.15810v1 Announce Type: new Abstract: Diffusion Transformers deliver state-of-the-art video generation, but their long spatiotemporal sequences make attention the dominant deployment cost,...
arXiv:2608. 12032v1 Announce Type: cross Abstract: Video diffusion transformers are costly to sample: every denoising step applies self-attention over a long 3D token sequence, a quadratic cost that dominates as resolution and duration grow.
arXiv:2610.29441v1 Announce Type: new Abstract: Full 3D attention is costly in video Diffusion Transformers. Block-sparse attention reduces this cost by computing only selected attention blocks. Howe...
arXiv:2609.37001v1 Announce Type: cross Abstract: Diffusion Transformers (DiTs) enable high-quality video generation but suffer from substantial inference latency, primarily attributable to the compu...
arXiv:2609.23153v1 Announce Type: new Abstract: Video diffusion transformers are expensive because attention dominates long spatiotemporal token sequences. We identify the \emph{high-sparsity trap}:...
PISA introduces a training‑free Piecewise Sparse Attention mechanism that preserves full attention span while reducing complexity. It keeps exact computation for critical blocks and approximates non‑critical ones using block‑wise Taylor expansion, avoiding the quality loss seen in conventional sparse attention. Experiments show significant speedups—up to 2.57× on Hunyuan‑Video—while maintaining or improving visual quality across video and image generation tasks.
TileMix is a tile‑centric mixed‑precision attention kernel that routes score‑tile groups within fused dense attention to either FP16 or INT8 computation, using compact bitmasks to decide precision per tile. By partitioning the attention matrix into hardware‑aligned tiles and updating a shared online‑softmax state, TileMix preserves dense token connectivity without requiring training and supports grouped‑query attention, variable‑length batches, and INT8 key/value caches. Benchmarks on LLaMA, Qwen, and Vicuna show that TileMix restores long‑context quality lost with uniform INT8 and improves prefill throughput over FP16, offering a controllable accuracy‑efficiency trade‑off across model families.
arXiv:2605.08371v2 Announce Type: replace Abstract: Multi-view geometry transformers are feed-forward 3D foundation models that jointly predict depth maps, point maps, and camera poses for N images i...
arXiv:2610.32233v1 Announce Type: new Abstract: Fixed-window Vision Transformer (ViT) decoders run at deployment resolution as overlapping tiles: every token in an overlap band is processed by severa...
arXiv:2610.07758v1 Announce Type: cross Abstract: Training-free token reduction accelerates vision transformers by removing redundant tokens across layers, recovering most of the original accuracy at...
arXiv:2609.15131v1 Announce Type: cross Abstract: Multimodal large language models (MLLMs) require substantial computation to process numerous visual tokens across all transformer layers. Most method...
arXiv:2603. 00040v3 Announce Type: replace-cross Abstract: Achieving reliable 4-bit attention is a prerequisite for end-to-end FP4 computation on emerging FP4-capable GPUs, yet attention remains the main obstacle due to FP4's tiny dynamic range and attention's heavy-tailed activations.