Skip to main content

ComfyUV Deployment Guide

Overview

This guide explains how to deploy the Docker image:

sandichhuu/comfyuv on Docker Hub

⚛️ ComfyUI - UV precompiled byte code.

An optimized, production-ready ComfyUI Docker image built with uv for ultra-fast Python package management. It comes pre-configured with modern performance optimizations like Triton, SageAttention.

⚠️ Important Note
🌵 I only have RTX 20, 30 series (SM75 & SM86). So newer series untested.
🌵 If you facing any issue, please send the error log to:
     💮 LinkedIn
     💮 Facebook
     💮 Whatsapp

🌵 Compatible GPU
✅ Newer : Runnable, but slower.
✅ SM120 RTX 50 Series. ✅ SM86 : RTX 40 Series.
✅ SM86 : RTX 30 Series.
✅ SM75 : RTX 20 Series.
⛔ Unsupported : GTX 10 Series and older, AMD or Intel GPU.

⚠️ Unsupport xformers.
      Because cu132 have no pre-build for this package & have no reason to use the slowest method.
✅ You can use sage-attention or triton instead.

⚠️ No longer support flash-attn (because the compile time is super slow).
✅ You can do it your self with this command:

MAX_JOBS=10 NVCC_THREADS=1 uv pip install flash-attn --no-build-isolation


🌸 Why use this Docker image for ComfyUI?

    ✅ Custom Nodes Compatibility: Python 3.13.13, widely compatible with extensions.


    ✅ Blazing Fast Execution: Leverages uv for ultra-fast package management and bytecode compilation. Built on CUDA 13.2, delivering superior performance compared to older versions like CUDA 12.8 or 13.0.
    ✅ Precompiled Ecosystem: Comes out-of-the-box with precompiled essential packages and popular custom nodes, saving you significant setup and compile time.
    ✅ Privacy-Focused (No Analytics): Completely stripped of telemetry and analytics tracking. Your data and workflows remain entirely private.
    ✅ Optimized for cloud: You can public the website without configuration. Use caddy or nginx for authentication.

    PRE-INSTALLED

      sage-attention

        🌸

        PRE-INSTALLED Custom Nodes
        ▫️ calcuis/gguf (calcuis/gguf) ▫️ BobRandomNumber/ComfyUI-Crystools-MonitorOnly (BobRandomNumber) ▫️ Kosinkadink/ComfyUI-VideoHelperSuite (Kosinkadink) ▫️ rgthree/rgthree-comfy (rgthree) ▫️ willmiao/ComfyUI-Lora-Manager (willmiao) ▫️ Comfy-Org/Nvidia_RTX_Nodes_ComfyUI (Comfy-Org) ▫️ ComfyUI-KJNodes (kijai)

        🌸 Quick Start (Docker Compose)

        Install Docker, Nvidia Driver (refer Studio version), CUDA Toolkit (version 13.2).
        Next, copy docker-compose.yml content below and put to empty folder.
        Then, open Terminal/CMD/Windows PowerShell and navigate to directory contain docker-compose.yml file.
        Finally, run this command docker compose up -d to start the container. Go to http://localhost:8188 and enjoy.

        Requirements

        ⬇️ Docker
        ⬇️ nvidia-driver
        ⬇️ cuda-toolkit (CUDA 13.2)

        docker-compose.yml

        services:
  comfyuv:
    image: sandichhuu/comfyuv:sm86cu132
    container_name: comfyuv
    ipc: host
    port:
      - 8188:8188
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ./input:/comfy/input
      - ./output:/comfy/output
      - ./workflows:/comfy/user/default/workflows
      - ./models:/comfy/models
      - ./workflows:/comfy/user/default/workflows
      
      - comfyuv:/comfy
      - uv_cache:/root/.cache/uv
    command: >
      --enable-triton-backend
#     --use-flash-attention unsupport becase it take long times to compile.
#     --lowvram
      --use-sage-attention
      --disable-pinned-memory
      --fast fp8_matrix_mult autotune
      --front-end-version Comfy-Org/ComfyUI_frontend@latest

volumes:
  comfyuv:
  uv_cache:

        🌸 Setup Structure

        comfyuv/
├── docker-compose.yml
├── input/
├── models/
├── workflows/
└── output/

        🌸 Why have no all-in-one image ?

        All GPU support int4, int8, fp16, fp32.
        But each GPU series has different physics structure due to different kernels and different attention optimization.

        For example RTX4090 support fp8 native, RTX5090 support fp8 & nvfp4 native.
        RTX40xx fastest with FlashAttention3, RTX50xx with FlashAttention4.
        If I compile everything on one image, the file size come super heavy.
        And the compile time is slow as hell.

        🌸 Why have no all-in-one image ?

        All GPU support int4, int8, fp16, fp32.
        But each GPU series has different physics structure due to different kernels and different attention optimization.

        For example RTX4090 support fp8 native, RTX5090 support fp8 & nvfp4 native.
        RTX40xx fastest with FlashAttention3, RTX50xx with FlashAttention4.
        If I compile everything on one image, the file size come super heavy.
        And the compile time is slow as hell.

        Back to top