Installation¶
Environment¶
conda create -n solengine python=3.12 -y
conda activate solengine
pip install torch==2.11.0 --index-url https://download.pytorch.org/whl/cu128
pip install diffusers==0.37.1 transformers accelerate imageio-ffmpeg
git clone -b sol-engine https://github.com/NVlabs/Sana.git
cd Sana
pip install ./techniques/sparse_backends
techniques/sparse_backends is the Sol-Attn kernel package. It needs PyTorch
already installed — it does not pull one in, so it cannot override a build
matched to your CUDA.
Sol-Attn ships CuTe kernels for sm90 (H100), sm100 (B200/GB200) and sm120
(RTX 5090), and a Triton reference elsewhere. benchmark.json records which
backend a run used.