This repository is dedicated to helping ComfyUI, Fooocus, FramePack users on Windows significantly enhance their AI workflow efficiency. It offers pre-compiled acceleration packages such as xformers, Flash Attention, and SageAttention, along with detailed installation guides.
If you need to upgrade to other versions of Nvidia graphics cards or a different version of Python or Torch, use the links to the official project pages
- Optimized Attention Packages: Directly downloadable, self-compiled versions of leading attention optimizers for ComfyUI, Fooocus, FramePack.
- xformers: A library providing highly optimized attention mechanisms.
- Flash Attention: Designed for ultra-fast attention computations.
- SageAttention: Another powerful tool for accelerating attention.
- Step-by-Step Installation Guides: Clear and concise instructions to seamlessly integrate these packages into your ComfyUI environment on Windows.
- Direct Download Links: Convenient links to quickly access the compiled files.
- GPU: All provided builds are specifically compiled and optimized for NVIDIA RTX 5000 series GPUs (Blackwell architecture). They are targeted at SM120 (Streaming Multiprocessor 120). Compatibility with other NVIDIA series is not guaranteed.
- Operating System: Windows.
- Python: The compilation process used Python 3.12.10. For optimal compatibility, it's recommended to use Python 3.12.x.
- Compilation Tools: The builds were created using CUDA Toolkit 12.8 and Visual Studio 2022.
If you're running ComfyUI on Windows with an NVIDIA RTX 5000 series GPU and looking to drastically reduce generation times and improve overall performance, this repository provides the necessary tools and guidance to get you started quickly.
There are generally three straightforward ways to deploy applications like ComfyUI, Fooocus, and FramePack. While their setup processes are quite similar, a common challenge arises with their bundled Python environments: cannot change the version of Python in their `python_embedded' folders.
Method 1: Portable Version (Download & Unpack): This method involves downloading and extracting a portable version of these applications. Each portable version comes with its own embedded Python environment, which might not always align with Python 3.12.x and would require installing accelerators from their respective repositories.
- Fooocus: Typically includes Python 3.10.9 and Torch 2.1.0+cu121. This version Torch does not work with the NVIDIA RTX 5000 series.
- FramePack: Usually comes with Python 3.10.6 and Torch 2.6.0+cu126. This version Torch does not work with the NVIDIA RTX 5000 series.
- ComfyUI: The latest version, v0.3.44, includes Python 3.12.10 and Torch 2.7.1+cu128, which fully meets our requirements for NVIDIA RTX 5000 series compatibility.
To ensure Fooocus and FramePack are compatible with the NVIDIA RTX 5000 series GPUs, you'll need to update their Torch libraries. According to the official PyTorch website, it supports Python versions 3.9 and higher, making it compatible with the embedded Python versions in these applications.
First, you need to open a command prompt (cmd) directly within the python_embedded folder of each application.
- For Fooocus: Navigate to this path:
X:\Fooocus_win64_2-5-0\python_embeded\ - For FramePack: Navigate to this path:
X:\framepack_cu126_torch26\system\python\
To open the console in these folders:
- Open File Explorer (Windows Explorer).
- Navigate to the specific
python_embededorpythonfolder for your application. - In the address bar at the top of the File Explorer window, type
cmdand press Enter. This will open a command prompt window directly in that directory.
For the 5000 series, you need to install Torch version 2.7.0 or higher, compiled with CUDA 128. Once the console is open in the correct folder, execute the following two commands in sequence:
-
Upgrade pip (the package installer):
.\python.exe -m pip install --upgrade pipThis ensures you have the latest version of pip, which is necessary for reliable package installations.
-
Upgrade Torch, torchvision, and torchaudio for CUDA 12.8:
.\python.exe -m pip install --upgrade torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128This command will download and install the latest versions (currently it is 2.7.1, to install another version, you need to register its revision torch==2.7.0) of Torch, torchvision (for computer vision utilities), and torchaudio (for audio utilities) that are compiled for CUDA 12.8. The
--extra-index-urlspecifies the official PyTorch wheel index for CUDA 12.8 builds, ensuring you get the correct versions for your RTX 5000 series GPU.
After successfully running these commands, your Fooocus and FramePack installations should have updated Torch versions, enabling proper functionality with your NVIDIA RTX 5000 series graphics card.
We go to method 2 and 3.
Before proceeding with either method, you need to install the core tools:
- Install Git: Download and install Git from its official website.
- Install Python 3.12.10: Download and install Python 3.12.10 from python.org. Crucially, remember to check the "Add Python to PATH" option during installation. This ensures Python is accessible from your command line.
This method involves using a globally installed Python (e.g., Python 3.12.10) for all your applications.
- How it works: When you install packages, they are placed in the global Python environment and become accessible to any application configured to use that Python installation.
- Pros: Simple to set up initially, as you don't manage separate environments.
- Cons: If you have multiple AI applications installed (like Fooocus, FramePack, and ComfyUI), they might require different versions of the same packages. This can easily lead to package conflicts, causing one or more applications to break. This approach is generally not recommended if you plan to run multiple Python-based applications simultaneously. If you only run one application, it might be acceptable.
This is the recommended approach for managing multiple AI applications. It involves using a virtual environment to isolate each application's dependencies, even while using a single global Python version.
- How it works: A virtual environment creates an isolated copy of a Python environment for each project. All packages installed within this virtual environment are specific to that project and won't interfere with other applications or your global Python installation. This ensures each application runs with its required package versions, specified in its
requirements.txt. - Pros: Prevents package conflicts between different applications, making your development and deployment more stable and reliable. Each application gets its own clean set of dependencies.
- Cons: Requires an extra step to create and activate the virtual environment for each project.
- Create a Project Folder:
First, create a new folder for your application (e.g.,
Fooocus_Git,FramePack_Git,ComfyUI_Git). It's best to place this folder in the root of a fast drive (e.g.,D:\ComfyUI_Git). - Clone the Application Repository:
Open the console (cmd) in the newly created application folder (for example, X:\ComfyUI_Git), and then clone the desired application from its official GitHub repository:
- For Fooocus:
git clone https://github.com/lllyasviel/Fooocus.git
- For FramePack:
git clone https://github.com/lllyasviel/FramePack.git
- For ComfyUI:
git clone https://github.com/comfyanonymous/ComfyUI.git
- For Fooocus:
- Create and Activate Virtual Environment:
Then, create and activate your virtual environment:
- Create the virtual environment:
python -m venv venv
- Activate it:
.\venv\Scripts\activate
(venv)before the drive letter and path, like this:This indicates that all subsequent package installations will be isolated to this virtual environment, specifically for the(venv) X:\ComfyUI_Git>ComfyUIapplication in this example. The same logic applies to other applications like Fooocus or FramePack, with their own dedicated virtual environments. - Create the virtual environment:
In the portable version, after unpacking, there are *.bat files for launching programs. When cloning, these are absent, so you need to create your own run.bat file next to the venv folder and the cloned program's folder, in the following format:
Fooocus:
@echo off
call .\venv\Scripts\activate.bat
python .\Fooocus\entry_with_update.py --theme dark
pauseFramePack:
@echo off
call .\venv\Scripts\activate.bat
python .\FramePack\demo_gradio_f1.py --server 127.0.0.1 --inbrowser
pauseComfyUI:
@echo off
call .\venv\Scripts\activate.bat
python .\ComfyUI\main.py --auto-launch
pauseYou can add all other launch arguments yourself according to your needs.
To install xformers for Fooocus, FramePack, or ComfyUI, execute the following commands in the console, depending on your version:
Open the console in your program's python_embedded folder and execute:
.\python.exe -m pip install -U xformers --index-url https://download.pytorch.org/whl/cu128Ensure you have activated your virtual environment, then execute:
pip install -U xformers --index-url https://download.pytorch.org/whl/cu128- If your Torch version is 2.7.0, install
xformerswith a specific version:pip install -U xformers==0.0.30 --index-url https://download.pytorch.org/whl/cu128
- For Torch 2.7.1 and higher, you can omit the
xformersversion (as shown in the commands above), as a compatible latest version will be installed. - To download my version of `xformers', make sure you get the link from the folder corresponding to your version of Torch.
Note: Fooocus does not support FlashAttention or SageAttention, so these instructions are only for FramePack and ComfyUI.
To install FlashAttention for FramePack or ComfyUI, execute the following commands in the console:
.\python.exe -m pip install https://huggingface.co/lldacing/flash-attention-windows-wheel/resolve/main/flash_attn-2.7.4.post1%2Bcu128torch2.7.0cxx11abiFALSE-cp312-cp312-win_amd64.whlpip install https://huggingface.co/lldacing/flash-attention-windows-wheel/resolve/main/flash_attn-2.7.4.post1%2Bcu128torch2.7.0cxx11abiFALSE-cp312-cp312-win_amd64.whl- This version works with both Torch 2.7.0 and 2.7.1.
- To download my version of `FlashAttention', make sure you get the link from the folder corresponding to your version of Torch.
Before installing SageAttention, you must install Triton. The aim of Triton is to provide an open-source environment to write fast code with higher productivity than CUDA, but also with higher flexibility than other existing DSLs.
.\python.exe -m pip install -U triton-windowspip install -U triton-windowsTo install SageAttention for FramePack or ComfyUI, execute the following commands in the console:
.\python.exe -m pip install https://github.com/woct0rdho/SageAttention/releases/download/v2.2.0-windows/sageattention-2.2.0+cu128torch2.7.1-cp312-cp312-win_amd64.whlpip install https://github.com/woct0rdho/SageAttention/releases/download/v2.2.0-windows/sageattention-2.2.0+cu128torch2.7.1-cp312-cp312-win_amd64.whl- If you need a different version, you can find it in the list here:
- To download my version of `SageAttention', make sure you get the link from the folder corresponding to your version of Torch.
This section details how different attention mechanisms are handled by each application and any specific flags you might need to use in their respective run.bat files.
Fooocus automatically works with an installed xformers library. You don't need any special flags to enable it. However, if you wish to disable xformers, you can do so by adding the following flag to your run.bat file:
--disable-xformersFramePack also automatically utilizes installed xformers, FlashAttention, and SageAttention. No specific flags are required in your run.bat file for these to function.
ComfyUI automatically works with xformers when installed. If you need to disable xformers for ComfyUI, add this flag to its run.bat file:
--disable-xformersFor FlashAttention, you need to explicitly activate it in ComfyUI's run.bat file. Crucially, if you use FlashAttention, you must disable xformers, as xformers has its own built-in FlashAttention code, and they are not designed to work concurrently.
To activate FlashAttention:
--use-flash-attentionConversely, SageAttention can operate in parallel with xformers in ComfyUI. To activate SageAttention, add this to your run.bat file:
--use-sage-attentionThe bench folder contains three tests that allow you to verify the functionality of all installed components:
environment.py: Displays information about your system and installed packages.fp32-16-8-4_bf16_gpt-bench.py: Runs a system benchmark with various parameters for Torch, xformers, and FlashAttention.sa-fa_gpt.py: Compares SageAttention and FlashAttention where applicable.