Extensions/ComfyUI-AutoMultiGPU
ComfyUI Extension

ComfyUI-AutoMultiGPU

A ComfyUI extension with 11 custom nodes.

By nexusfinancial-dev·Created 4 days ago·Updated 4 days ago· 1
nexusfinancial-dev/ComfyUI-AutoMultiGPU
Nodes11
On cloudLocal install
CategoryAutoMultiGPU, AutoMultiGPU/modular
Stars1
Updated4 days ago

Nodes (11)

Auto Checkpoint Loader (MultiGPU)
AutoMultiGPU
Auto CLIP Loader (MultiGPU - Safetensors & GGUF)
AutoMultiGPU/modular
Auto Device Router (MultiGPU)
AutoMultiGPU/utils
Auto DualCLIP Loader (MultiGPU - Safetensors & GGUF)
AutoMultiGPU/modular
Auto MiniMax H3 VAE Decode (Zero-OOM)
AutoMultiGPU/minimax_h3
AutoMultiGPU - Automatic Engine (Zero-OOM Governor)
AutoMultiGPU/Engines
AutoMultiGPU - DisTorch Engine (Layer Sharding)
AutoMultiGPU/Engines
AutoMultiGPU - PyTorch Engine (Native Device Split)
AutoMultiGPU/Engines
Auto UNET Loader (MultiGPU - Safetensors & GGUF)
AutoMultiGPU/modular
Auto Universal Loader (MultiGPU - Safetensors & GGUF)
AutoMultiGPU
Auto VAE Loader (MultiGPU)
AutoMultiGPU/modular
Readme

⚡ ComfyUI-AutoMultiGPU

🚀 High-Performance Zero-OOM Multi-GPU Distribution & Memory Management Engine for ComfyUI

License: MIT PyTorch 2.4+ CUDA 12.x+ ComfyUI Compatible GitHub Stars

ComfyUI-AutoMultiGPU is a production multi-GPU engine designed to run massive generative AI models (FLUX.2 / Klein 9B, LTX-Video 2.5, SDXL, Wan2.1, MiniMax H3, and GGUF quants) across dual or multi-GPU configurations with Zero Out-Of-Memory (OOM) errors.

🏛️ Core Engines

1. ⚙️ AutoMultiGPU - PyTorch Engine (Native Device Split)

Splits pipeline components natively across separate physical CUDA devices:

  • Primary GPU (cuda:0): Dedicated exclusively to UNet / Diffusion Transformer inference.
  • Secondary GPU (cuda:1): Hosts Text Encoders (CLIP, T5, Qwen3), VAE Encoders/Decoders, and ControlNets.
  • Zero-OOM CPU Offloading: Dynamically swaps non-active weights to system RAM.

2. ⚡ AutoMultiGPU - DisTorch Engine (Layer Sharding)

  • Shards individual Transformer blocks of large DiT models (e.g. Klein 9B / FLUX) across GPUs:
    • GPU 0: Layers 0 → 18
    • GPU 1: Layers 19 → 37
  • Provides massive speedup by utilizing combined VRAM without tensor duplication.

3. 🧠 AutoMultiGPU - Automatic Engine (Zero-OOM Governor)

  • Dynamically detects available VRAM across all installed NVIDIA GPUs at runtime and balances workloads automatically.

📊 Benchmark Comparison (Dual RTX 5060 Ti 16GB)

| Model & Workflow | Single GPU (16GB) | AutoMultiGPU Engine (2x 16GB) | Speedup & Stability | | :--- | :--- | :--- | :--- | | FLUX.2 Klein 9B (5-Ref Scene) | ⚠️ OOM / Crash | ✅ 0.42s per step (Zero-OOM) | 100% Stable | | LTX-Video 2.5 (High Res Take) | ⚠️ OOM at Frame 49 | ✅ 121 Frames Generated | Zero VRAM Spills | | 4K Super-Resolution Refiner | ⚠️ Swapping to CPU (Slow) | ✅ Full GPU Acceleration | 3.8x Faster |

📥 Installation

Method 1: Via ComfyUI Manager (Recommended)

  1. Open ComfyUI Manager.
  2. Search for ComfyUI-AutoMultiGPU.
  3. Click Install and restart ComfyUI.

Method 2: Manual Git Clone

cd ComfyUI/custom_nodes
git clone https://github.com/nexusfinancial-dev/ComfyUI-AutoMultiGPU.git
cd ComfyUI-AutoMultiGPU
pip install -r requirements.txt

📂 Example Workflows

Sample workflows are available in the examples/ folder. Drag and drop any .json file directly into ComfyUI!

📜 License

This project is open-source and licensed under the MIT License. Created with ❤️ by nexusfinancial-dev.