Wan2.2

Headlines · 2

1. Oct 20
   ALLOT: training-free block-sparse attention allocation for video DiTs

   A new arXiv preprint by Tianze Xia and five co-authors proposes ALLOT (Allocating Layer-step Latency by Output Tests), a training-free method for allocating block-sparse attention in video Diffusion Transformers. The authors observe that sparsifying attention in different layers and denoising steps affects final-video fidelity to very different degrees, while the most sensitive layer-step pairs stay nearly the same across inputs. ALLOT probes each layer-step cell and calibrates along the denoising trajectory, then jointly optimizes cell densities within a measured latency budget; the paper reports an 11.8 dB PSNR gain over XAttention at equal or shorter wall-clock time on HunyuanVideo, and wins or ties the default in all nine settings across Wan2.1, Wan2.2 and HunyuanVideo.

   Hugging Face · Papers · 🔥 0
2. Oct 20
   PyTorch 2.15 MPS adds FP8 and faster LLM decoding on Macs

   Hugging Face's PyTorch team published a blog post on October 20, 2026 describing MPS improvements in PyTorch 2.15, including FP8 support with float8_e4m3fn storage and scaled matrix multiplication on the Mac backend. The team reports self-measured gains: MoE decoding 5–6.5× faster than 2.14, with Qwen3.6-35B-A3B rising from 9.4 to 61.3 tokens/s on an M3 Ultra, and Wan2.2 video VAE decode peak memory falling from 270 GiB to about 8 GiB. The post also says MPS now requires macOS 15 or later, PyTorch 2.15 requires Python 3.11 or later, and 2.16 will run attention entirely on Metal.

   Hugging Face · 🔥 0

Experience and discussion from the community

Share my Wan2.2 experienceAsk about Wan2.2