mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-04-15 04:52:31 +08:00

Restore SageAttention, fix gpus syntax, and clean up entrypoint

SageAttention research confirms it remains useful (2-5x speedup over FA,
active development through SA3, broad community adoption for video/HR).
Restore it as an opt-in startup-compiled feature (FORCE_SAGE_ATTENTION=1).

Entrypoint is cleaned up vs the original: simplified GPU probe (drops
per-arch flag exports, keeps what's needed for strategy selection),
cleaner build logic with merged clone/update paths, removed dead code.

- Dockerfile: restore SAGE_ATTENTION_AVAILABLE=0 default env var
- entrypoint.sh: simplified but functionally equivalent SageAttention
  support; removes ~150 lines of redundant per-arch flag tracking
- README: fix docker-compose gpus syntax to 'gpus: all'; restore
  SageAttention docs with accurate description of behavior

https://claude.ai/code/session_01WQc56fWdK329K11kRGnb5g

2026-03-27 12:40:33 +00:00

5.0 KiB

Raw Blame History

ComfyUI-Docker

An automated Repo for ComfyUI Docker image builds, optimized for NVIDIA GPUs.

About • Features • Getting Started • Usage • License

About

This image packages upstream ComfyUI with CUDA-enabled PyTorch and an entrypoint that handles volume permissions and custom node setup.

The base image is python:3.12-slim (Debian trixie) with CUDA 12.8 developer libraries installed via apt and PyTorch installed from the cu128 wheel index.

It syncs with the upstream ComfyUI repository, builds a Docker image on new releases, and pushes it to GitHub Container Registry (GHCR).

I created this repo for myself as a simple way to stay up to date with the latest ComfyUI versions while having an easy-to-use Docker image.

Features

Daily checks for upstream releases, auto-merges changes, and builds/pushes Docker images.
CUDA-enabled PyTorch + Triton on Debian trixie with CUDA 12.8 dev libs so custom CUDA builds work at runtime.
Non-root runtime with PUID/PGID mapping handled by entrypoint for volume permissions.
ComfyUI-Manager auto-sync on startup; entrypoint scans custom_nodes and installs requirements when COMFY_AUTO_INSTALL=1.
SageAttention build-on-start for compatible NVIDIA GPUs (Turing/SM 7.5+); enabling is opt-in via FORCE_SAGE_ATTENTION=1.

Getting Started

Install NVIDIA Container Toolkit on the host, then use docker run --gpus all or Compose GPU reservations to pass GPUs through.
Expose the ComfyUI server on port 8188 (default) and map volumes for models, inputs, outputs, and custom_nodes.

Pulling the Image

The latest image is available on GHCR:

docker pull ghcr.io/clsferguson/comfyui-docker:latest

For a specific version (synced with upstream tags, starting at v0.3.59):

docker pull ghcr.io/clsferguson/comfyui-docker:vX.Y.Z

Docker Compose

For easier management, use this docker-compose.yml:

services:
  comfyui:
    image: ghcr.io/clsferguson/comfyui-docker:latest
    container_name: ComfyUI
    runtime: nvidia
    restart: unless-stopped
    ports:
      - 8188:8188
    environment:
      - TZ=America/Edmonton
      - PUID=1000
      - PGID=1000
    gpus: all
    volumes:
      - comfyui_data:/app/ComfyUI/user/default
      - comfyui_nodes:/app/ComfyUI/custom_nodes
      - /mnt/comfyui/models:/app/ComfyUI/models
      - /mnt/comfyui/input:/app/ComfyUI/input
      - /mnt/comfyui/output:/app/ComfyUI/output

Run with docker compose up -d.

Usage

Open http://localhost:8188 after the container is up; change the external port via -p HOST:8188.
To target specific GPUs, use Docker's GPU device selections or Compose device_ids in reservations.

SageAttention

SageAttention is compiled at container startup when a compatible GPU (Turing SM 7.5+) is detected and cached to a volume-mapped directory for subsequent starts. It delivers 2-5x faster attention vs FlashAttention for video and high-res image workflows.

To enable: set FORCE_SAGE_ATTENTION=1. If the build or import fails, ComfyUI starts normally without it.
The first startup with SageAttention will be slower due to compilation; subsequent starts use the cached build.
Turing GPUs (RTX 20xx) use the v1.0 branch with Triton 3.2.0; Ampere and newer use the latest release.

Environment Variables

PUID/PGID: map container user to host UID/GID for volume write access.
COMFY_AUTO_INSTALL=1: auto-install Python requirements from custom_nodes on startup (default: 1).
COMFY_FORCE_INSTALL=1: force reinstall of custom_nodes requirements even after first run.
FORCE_SAGE_ATTENTION=0|1: compile and enable SageAttention on startup (requires compatible NVIDIA GPU).
SAGE_MAX_JOBS=N: override the number of parallel compile jobs for SageAttention (default: auto from RAM).
CM_*: seed ComfyUI-Manager config.ini keys on first start (e.g. CM_SKIP_UPDATE_CHECK=1).

License

Distributed under the MIT License (same as upstream ComfyUI). See LICENSE for more information.

Contact

Creator: clsferguson - GitHub
Project Link: https://github.com/clsferguson/ComfyUI-Docker

Built with ❤️ for easy AI workflows.

5.0 KiB Raw Blame History