The first video is sage-attention 2.2 https://github.com/user-attachments/assets/19fea213-93cf-4b2b-b44a-ff4f5a37dfc2 The second video is sage-attention 3 https://github.com/user-attachments/assets/038e4539-2ef9-4326-9ce5-c37abc67be21 Both videos take about same time to generate, however sage-attention 3 produces mosaic artifacts, whereas sage-attention 2.2 is very clean. I am using DGX Spark, Python 3.11, https://download.pytorch.org/whl/cu130 and ``` # sage-attention 2.2 python setup.py bdist_wheel # sage-attention 3 ( cd sageattention3_blackwell python setup.py bdist_wheel ) # sage-attention 2.2 pip install ~/SageAttention/dist/sageattention-2.2.0-cp312-cp312-linux_aarch64.whl # Sage-attention 3 pip install ~/SageAttention/sageattention3_blackwell/dist/sageattn3-1.0.0-cp312-cp312-linux_aarch64.whl ```