Attention Kernel Used in SGLang diffusion
VMoBA: Mixture-of-Block Attention for Video Diffusion Models (VMoBA)
Installation
Please ensure that you have installed FlashAttention version 2.7.1 or higher, as some interfaces have changed in recent releases.
Usage
You can use moba_attn_varlen in the following ways:
Install from source:
python setup.py install
Import after installation:
from vmoba import moba_attn_varlen
Or import directly from the project root:
from csrc.attn.vmoba_attn.vmoba import moba_attn_varlen
Verify if you have successfully installed
python csrc/attn/vmoba_attn/vmoba/vmoba.py