SageAttention on NVIDIA RTX 50-series (Blackwell) - 30% Faster, Native Windows

[mobcat40]

SageAttention + PyTorch 2.11 Nightly on RTX 50-series (Blackwell) - 30% Faster, Native Windows

Got SageAttention 2.2.0 working on Blackwell GPUs with native Windows - no Docker or WSL needed. ~30% faster diffusion sampling.

Download

Prebuilt wheel: https://github.com/mobcat40/sageattention-blackwell

Requirements

• RTX 5090/5080/5070 (Blackwell, sm_120)
• PyTorch 2.11 nightly + CUDA 12.8
• Python 3.11
• Windows 10/11

Install (2 minutes)

# For ComfyUI portable:
.\python_embeded\python.exe -m pip install sageattention-2.2.0+cu128.torch2.11-cp311-cp311-win_amd64.whl

# For venv:
path\to\venv\Scripts\python.exe -m pip install sageattention-2.2.0+cu128.torch2.11-cp311-cp311-win_amd64.whl

Usage in ComfyUI

Don't use --use-sage-attention flag - it uses the Triton backend which causes black output with some models (Qwen, Wan).

Instead:

1. Install ComfyUI-KJNodes
2. Add "Patch Sage Attention" node to your workflow
3. Set backend to sageattn_qk_int8_pv_fp16_cuda

Performance

~30-35% faster on RTX 5090 (tested with Qwen Image Edit: 14m30s → 9m30s for 40 steps)

Why this exists

• pip install sageattention installs 1.0.6 which causes black output
• woct0rdho wheels are built for PyTorch 2.9, DLL load error on 2.11
• Building from source hits MSVC "ambiguous symbol 'std'" error

Full build instructions if you want to compile yourself: https://github.com/mobcat40/sageattention-blackwell/blob/main/BUILD_STORY.md

[picolo100]

what about linux is this working on it, seeking sageattention 2.2 working on rtx 6k pro, plz lemme know, urgennnnnnnnnnnnnt

[mobcat40]

The wheels I posted are Windows-only (.whl for win_amd64).

For Linux, building from source should be easier - you won't hit the MSVC/VS2025 issues. You may still need to patch the PyTorch header if you're on 2.11 nightly (the std ambiguity bug in compiled_autograd.h).

What GPU exactly? "RTX 6K Pro" - do you mean RTX 6000 Ada (Lovelace)? That's sm_89, not Blackwell, and should work with standard PyTorch releases + official SageAttention wheels. You'd only need my build process if you're on an actual RTX 50-series (Blackwell/sm_120).

[picolo100]

it's this specs :
may i have you on telegram for faster interactions ?
NVIDIA-SMI 590.44.01 Driver Version: 590.44.01 CUDA Version: 13.1
NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 12.0

[ChiefNakor]

Try this @picolo100
https://github.com/ChiefNakor/comfyui-blackwell-docker