ComfyUI Triton and SageAttention Installer

Cross-platform installer for Triton and SageAttention on ComfyUI. Simplifies GPU-accelerated inference setup for Windows users with automated dependency management and RTX 5090 support.

Overview

Installing SageAttention on Windows has been notoriously difficult due to compilation issues, missing dependencies, and platform-specific challenges. This installer automates the entire process, making advanced GPU optimizations accessible to all ComfyUI users.

Features

One-command installation - No manual compilation or complex setup required
Windows-first design - Specifically addresses Windows compilation challenges
Automatic CUDA detection - Detects your GPU and CUDA version automatically
Pre-built wheel support - Uses pre-compiled wheels to avoid build issues
Visual Studio automation - Handles Build Tools installation if needed
Python header management - Automatically sets up development headers
RTX 5090 ready - Full support for Blackwell architecture with CUDA 12.8
Graceful fallbacks - Continues installation even if some components fail
Detailed logging - Comprehensive logs for troubleshooting

Prerequisites

Windows 10/11 (stubs for Linux/macOS experimental)
NVIDIA GPU with CUDA support
Existing ComfyUI installation
Python 3.8+ (or ComfyUI portable's embedded Python)

Installation

Quick Start

python comfyui_triton_sageattention.py --install --with-custom-nodes --backup

Note: Run from your ComfyUI directory by doing cd C:\path\to\ComfyUI, or use --base-path C:\path\to\ComfyUI to specify the location. If you're unsure where ComfyUI is installed, use --install discover to automatically find it.

By default comfyui_triton_sageattention.py installs Triton and SageAttention 2 (falls back to 1 if unavailable). Add --with-custom-nodes to also install VideoHelperSuite and DazzleNodes. It's recommended to add --backup to automatically save your environment first, in case you opt to restore later.

Find ComfyUI Installations

Not sure where ComfyUI is installed? The installer can find it for you:

# List all found ComfyUI installations python comfyui_triton_sageattention.py --show-installed locations Install with interactive discovery (finds and lets you select)

python comfyui_triton_sageattention.py --install discover

Check Current Installation

See what's already installed before or after running the installer:

python comfyui_triton_sageattention.py --show-installed

Preview Changes (Dry Run)

Preview what would be installed or upgraded without making any changes:

# Preview install
python comfyui_triton_sageattention.py --install --dryrun  #or --upgrade

Additional Installation Options

# Verbose mode for debugging python comfyui_triton_sageattention.py --install --verbose Custom ComfyUI location python comfyui_triton_sageattention.py --install --base-path C:\path\to\comfyui Cleanup previous installation

python comfyui_triton_sageattention.py --cleanup

ComfyUI Portable Support

For ComfyUI Portable distributions (the pre-packaged version with python_embeded folder):

# Run from your ComfyUI Portable directory cd C:\ComfyUI_windows_portable python path\to\comfyui_triton_sageattention.py --install # Or to run from any folder without "cd" simply add: --base-path C:\ComfyUI_windows_portable Check what the installer detected

python comfyui_triton_sageattention.py --show-installed --base-path C:\ComfyUI_windows_portable

The installer automatically detects the python_embeded folder and uses that Python environment instead of creating a new virtual environment. All packages are installed into the portable distribution's embedded Python.

Python Environment Selection

Control which Python environment the installer uses with the --python flag:

# Default --install auto-detects (portable > .venv > venv > system)
# To override, use --python with one of the options:
python comfyui_triton_sageattention.py --install --python {auto, system, venv, portable, .\path\...}
# --python also supports relative / absolute paths: .\path\to\venv, C:\Python312\python.exe

Note: To distinguish keywords from folder names, paths must include a separator (/, \, or start with ./). For example, --python venv uses the keyword (venv at base-path), while --python .\venv uses the folder in the current working directory.

For full parameter documentation, see docs/parameters.md.

SageAttention Version Control

Control which SageAttention version gets installed (default: SA 2.2.0.post3):

# Explicitly install SageAttention 1 (~2.1x speedup, more compatible) or 2 (~3x speedup) python comfyui_triton_sageattention.py --install --sage-version 1 # or 2 Opt-in to experimental/prerelease versions (use with caution) python comfyui_triton_sageattention.py --install --experimental Upgrade existing installation to latest version

python comfyui_triton_sageattention.py --upgrade

For a full list of supported PyTorch/CUDA/Python combinations, see docs/supported_wheels.md.

Environment Backup

Before making changes, it's helpful to create a backup of your environment:

# Create a backup before upgrading python comfyui_triton_sageattention.py --backup List available backups python comfyui_triton_sageattention.py --backup list Restore from the most recent backup (by index) python comfyui_triton_sageattention.py --backup-restore 1 Clean up old backups (keep only the latest)

python comfyui_triton_sageattention.py --backup-clean --keep-latest 1

Backups include:

Full copy of your venv or python_embeded folder
requirements.txt (pip freeze) for reference
RESTORE.txt with manual restore instructions

Usage

After installation, run ComfyUI with SageAttention enabled:

# Windows run_nvidia_gpu.bat Or manually

python ComfyUI\main.py --use-sage-attention

What Gets Installed

PyTorch with CUDA support matching your GPU
Triton-Windows - OpenAI's Triton for Windows
SageAttention - Efficient attention mechanism
Python Development Files - Required headers and libs

With `--with-custom-nodes`

ComfyUI-VideoHelperSuite - Video encoding/decoding utilities
DazzleNodes - DazzleML node collection

Troubleshooting

SageAttention Versions

The installer attempts to install SageAttention 2.x (faster, CUDA + Triton) when your PyTorch/CUDA/Python combination matches a pre-configured wheel. If no match is found, it falls back to SageAttention 1.0.6 (Triton-only).

Check your installed version:

python -m pip show sageattention

Version 1.0.6 = SageAttention 1 (~2.1x speedup vs FlashAttention2)
Version 2.x.x = SageAttention 2 (~3x speedup vs FlashAttention2)

For detailed information about version differences and troubleshooting, see Discussion #8: Understanding SageAttention Versions.

Common Issues

"No module named 'torch'" during build

The installer uses pre-built wheels to avoid this
Falls back to compatible versions automatically

Compilation errors

Python development headers are downloaded automatically
Visual Studio Build Tools are installed if needed

CUDA version mismatch

Installer auto-detects your CUDA version
Update NVIDIA drivers if issues persist

Getting SageAttention 1.0.6 instead of 2.x

Your PyTorch/CUDA/Python combo may not have a pre-configured wheel
You can manually install from woct0rdho's releases
Please open an issue so we can add your configuration

Logs

Check comfyui_install.log for detailed information.

Contributing

Contributions are welcome! Feel free to submit a pull request.

Like the project?

Development Setup

git clone https://github.com/djdarcy/comfyui-triton-and-sageattention-installer.git
cd comfyui-triton-and-sageattention-installer
python comfyui_triton_sageattention.py --install --verbose

How It Works

Detects platform and creates appropriate handler
Installs Visual Studio Build Tools (Windows)
Detects CUDA version for compatibility
Installs PyTorch with matching CUDA support
Installs Triton (triton-windows on Windows)
Downloads Python development headers
Installs SageAttention (pre-built wheels preferred)
Clones required repositories
Creates launch scripts

Additional Tools

AI Toolkit RTX 5090 Support

This repository also includes an installer for Ostris AI Toolkit users with RTX 5090 GPUs. The AI Toolkit installer (ai_toolkit_rtx5090_installer.py) addresses Blackwell architecture compatibility issues for LoRA training:

# From your AI Toolkit directory
python ai_toolkit_rtx5090_installer.py --non-interactive

See docs/ai_toolkit_rtx5090_setup.md for details.

Acknowledgments

woct0rdho - Windows wheels and triton-windows
thu-ml - SageAttention project
ComfyUI - The amazing UI
Ostris - AI Toolkit for LoRA training
ComfyUI community - Testing and feedback

License

This project is licensed under the MIT License - see the LICENSE file for details.

NAME

SYNOPSIS

INFO

DESCRIPTION

README