ExLlama nodes for ComfyUI.

Installation

Clone or download the repository to custom_nodes in your ComfyUI directory:

git clone https://github.com/Zuellni/ComfyUI-ExLlama

Install the latest ExLlama package from https://github.com/jllllll/exllama/releases. Choose the version matching your platform, Python, and PyTorch CUDA/ROCm. Example for Windows with Python 3.10 and CUDA 11.7:

pip install https://github.com/jllllll/exllama/releases/download/0.0.17/exllama-0.0.17+cu117-cp310-cp310-win_amd64.whl

Nodes

ExLlama Loader

Used to load 4-bit GPTQ Llama/2 models. You can find a lot of them over at https://huggingface.co/TheBloke. Specify the model directory in model_dir. ExLlama allocates memory according to max_seq_len, setting it lower is a good way to save on VRAM. Currently it's not possible to offload the model to CPU RAM.

ExLlama Generator

Generates a string based on the given prompt and max tokens for use with other nodes. Default values correspond to the simple-1 preset from https://github.com/oobabooga/text-generation-webui. ExLlama isn't deterministic so the outputs may differ even with the same seed.

Workflow

The example above can be opened in ComfyUI. Model used: https://huggingface.co/TheBloke/MythoLogic-Mini-7B-GPTQ. Total VRAM usage hovering just below 5GB without any SD model loaded.

Some random outputs:

• A vibrant underwater cityscape filled with colorful marine life, surrounded by sunken ruins from ancient civilizations.
• A romantic scene between two lovers during sunset, with flowers around them and birds flying overhead.
• A steampunk-inspired portrait of Jane Austen, dressed in Victorian era clothing and holding a parasol.