Modelpublic

sntdismas/Sageattention-3-torch2.11.0-cu130-py312-Blackwell

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes

Model Card

SageAttention 3 Wheel for Blackwell / CUDA 13.0 / PyTorch 2.11

This repository contains a locally compiled Python wheel for SageAttention 3 built for NVIDIA Blackwell GPUs.

The wheel was compiled for a RunPod ComfyUI environment using:

—Python: 3.12
—PyTorch: 2.11.0+cu130
—CUDA: 13.0
—GPU architecture: Blackwell sm_120
—Target platform: Linux x86_64
—Package: sageattn3

About SageAttention 3

SageAttention 3 is a Blackwell-oriented attention implementation designed to take advantage of NVIDIA RTX 50-series / Blackwell capabilities, including NVFP4-related acceleration paths.

It is intended for experimental acceleration of heavy image/video generation workloads such as ComfyUI pipelines with models like Flux or Wan, where attention performance can be a major bottleneck.

Source and credits

This wheel was built using the official SageAttention source repository:

—https://github.com/thu-ml/SageAttention

The build process was based on the RunPod compilation guide published by Seryoger on HuggingFace:

—https://huggingface.co/Seryoger/Sageattention-3-cu130-5090-endpoint

This wheel was rebuilt locally for a newer ComfyUI environment using:

—PyTorch 2.11.0+cu130
—CUDA 13.0
—Blackwell `sm_120`

Build notes

The wheel was compiled inside a RunPod Blackwell pod.

The important build parameters were:

bash

export TORCH_CUDA_ARCH_LIST="12.0"