How to Run Micromamba in ComfyUI with Triton, Sage-Attention, Flash-Attention, and X-formers for ComfyUI 3D

In this guide, I’ll walk you through setting up Micromamba to integrate Triton, Sage-Attention, Flash-Attention, and X-formers for ComfyUI 3D. This combination ensures smooth performance for advanced workflows, especially for 3D models and optimizations.

ComfyUI

git clone https://github.com/comfyanonymous/ComfyUI.git

A. Download and Configure Micromamba Directory

1. Open PowerShell as Administrator
   Not your regular PowerShell—make sure it’s Administrator-level access.
2. Create a Directory for Micromamba
   Choose where you want Micromamba to live. Examples:

   mkdir \micromamba

   or

   mkdir ~\micromamba

   Navigate to the new directory:

   cd \micromamba

3. Download the Latest Micromamba Windows Binary
   Use this command to grab the latest version:

   Invoke-WebRequest -URI https://micro.mamba.pm/api/micromamba/win-64/latest -OutFile micromamba.tar.bz2

4. Extract Micromamba
   Use 7-zip or your preferred unzip tool to extract:

   7z e micromamba.tar.bz2 && 7z x micromamba.tar && MOVE -Force Library\bin\micromamba.exe micromamba.exe

• Once extracted, you can safely delete the leftover files:

  • Library

    folder

  • .tar

    and

    .tar.bz2

    files
• Close PowerShell when you’re done.

C. Configure Environment Variables

1. Create the

   MAMBA_ROOT_PREFIX

   Environment Variable
   • Open Control Panel -> Edit System Environment Variables -> Environment Variables
2. Add the Micromamba Directory to Your PATH
   • Point to the folder where Micromamba lives (e.g.,

     C:\Users\username\AppData\Roaming\mamba

     ).

Open a new PowerShell session and check that Micromamba is installed properly:

micromamba -v

If it spits out a version number, you’re good to go!

Micromamba is lightweight, fast, and great for automating Python environment setups—especially when sharing with team members. Less headache, more coding.

1. Open an Administrator PowerShell window.
2. Type this command:

   Set-ExecutionPolicy Unrestricted

3. When prompted, press A (Yes to All).
4. Close the PowerShell window.

If you’re using a “debloated” version of Windows and PowerShell isn’t cooperating, install WezTerm as an alternative terminal:

• Download WezTerm here: WezTerm Install
• Or use this command via Winget:

  winget install wez.wezterm

To avoid issues later, create a micromamba environment. If you’re not familiar with micromamba, here’s how:

1. Install Micromamba:
   • Download it here: Micromamba Installation
2. Open your terminal and run:

   micromamba install xtensor -c conda-forge

3. Create a clean environment for ComfyUI:

   micromamba create -n 3d12 xtensor -c conda-forge

4.  micromamba activate 3d12

5. Replace

   3d12

   with any name you like (short and sweet, like

   ai

   or

   comfy

   ).
6. Install Python 3.12.1:

   micromamba install -c conda-forge python=3.12.1

7. Later, when installing ComfyUI dependencies, use:

   pip install -r requirements.txt

critical libmamba Shell not initialized

‘micromamba’ is running as a subprocess and can’t modify the parent shell.
Thus you must initialize your shell before using activate and deactivate.

To initialize the current powershell shell, run:
$ micromamba.exe shell hook -s powershell | Out-String | Invoke-Expression
and then activate or deactivate with:
$ micromamba activate
To automatically initialize all future (powershell) shells, run:
$ micromamba shell init –shell powershell –root-prefix=~/.local/share/mamba
If your shell was already initialized, reinitialize your shell with:
$ micromamba shell reinit –shell powershell
Otherwise, this may be an issue. In the meantime you can run commands. See:
$ micromamba run –help

Supported shells are {bash, zsh, csh, posix, xonsh, cmd.exe, powershell, fish, nu}.

micromamba.exe shell hook -s powershell | Out-String | Invoke-Expression

1. Download and install Git:
   • Git Download
2. Install Git LFS:
   • Git LFS Install

Make sure Git is on your system PATH. Run this to check:

$env:Path += ";C:\Program Files\Git\cmd"

ComfyUI Install

Git clone https://github.com/comfyanonymous/ComfyUI.git

1. Download the CUDA Toolkit 12.4:
   • CUDA 12.4 Download
2. Remove any older CUDA versions from your PATH to avoid conflicts.
3. Restart your computer after installation.

Important: Double-check that CUDA is correctly set up by ensuring its folder path exists in your system environment variables.

Xformers speeds up image generation. Install it before other dependencies:

1. Run this command to install Xformers with PyTorch:

   pip install -U xformers torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

   Replace

   cu124

   if you’re using a different CUDA version.

Need more info? Check out the official guides:

• Xformers Guide
• PyTorch Install

1. Download Visual Studio Build Tools here:
   • VS Build Tools
2. During installation, select:
   • Desktop Development with C++
   • Windows 11 SDK (10.0.22621.0 or newer)
   • MSVC v143 – VS 2022 (x64/x86)
   • C++ CMake Tools
   • C++ ATL headers and libraries

Time to fix those environment paths:

1. Open Windows Settings.
2. Search for “Edit system environment variables”.
3. Click on Environment Variables.
4. Add or edit these variables:
   • CUDA Path
   • Python PATH (add

     venv/Scripts

     folder).
   • Git Path

     C:\Program Files\Git\cmd

   • Visual Studio Tools Path (important for Triton):

     C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Auxiliary\Build

If you’re missing vcruntime files, here’s how to fix it:

1. Download the runtime installer:
   • VC Redist x64
2. Copy

   vcruntime140.dll

   and

   vcruntime140_1.dll

   into the Scripts folder of your Python environment.

https://www.mediafire.com/file/rtr9twx3qpx96dc/vcruntime140_1.dll/file

https://www.mediafire.com/file/gfjvkquaxn7o91u/vcruntime140.dll/file

1. Download the correct Triton wheel for your Python version:
   • Triton Wheels
2. Place the wheel file in your working directory (or provide its path):

   pip install triton-3.0.0-cp312-cp312-win_amd64.whl

   For embedded Python:

   ./python.exe -s -m pip install triton-3.0.0-cp312-cp312-win_amd64.whl

1. Download the SageAttention script:
2. Place the script in this folder:

   ComfyUI/custom_nodes/

3. Install it:

   pip install sageattention

Flash Attention 2 can help with performance optimization.

1. Check if you have Ninja installed:

   ninja --version

2. If not, install it:

   pip install ninja

3. Clone the Flash Attention repo:

   git clone https://github.com/Dao-AILab/flash-attention 

4. cd flash-attention

5. Compile and install:

   set MAX_JOBS=4 

6. pip install flash-attn --no-build-isolation

7. Run

   pip list

   to confirm it installed successfully.

Phew! You’re all set now. Activate your ComfyUI environment and start creating:

micromamba activate 3d12  # Or your environment name

If you run into any snags, just double-checked the system paths and installed versions. Now go enjoy some super-smooth, high-speed image generation!