ComfyUI-LLM-GGUF

model:https://huggingface.co/eddy1111111/qwen3_coder_8b_COT_un/blob/main/Qwen3-8B-Coder-Abliterated-Q4_K_M.gguf

ComfyUI custom nodes for LLM GGUF model inference.

Features

• Load and cache GGUF LLM models
• Chat inference with conversation history (hot session)
• Session management (reset, list)
• Model cache management
• Output condition prefix (Chinese T5 format)
• Supports llama-cpp-python binding or subprocess fallback

Nodes

Load GGUF Model

Load a GGUF LLM model for inference.

Inputs:

• model_name: Select GGUF model file from models/LLM folder
• gpu_layers: Number of layers to offload to GPU (default: 99 = all)
• context_size: Context window size in tokens (default: 32768)
• llama_cli_path: (Optional) Path to llama-cli.exe for subprocess fallback

Output:

• model: LLM_MODEL object

LLM Chat

Run LLM inference with chat support.

Inputs:

• model: LLM model from LoadGGUFModel
• user_prompt: User message to send to the model
• system_prompt: (Optional) System prompt to set model behavior
• session_id: Session ID for maintaining conversation history (default: "default")
• max_tokens: Maximum tokens to generate (default: 1024)
• temperature: Sampling temperature (default: 0.7)
• top_p: Top-p sampling threshold (default: 0.9)
• top_k: Top-k sampling (default: 40)
• repeat_penalty: Repetition penalty (default: 1.1)
• keep_history: Keep conversation history for hot session (default: True)
• output_condition: Output condition prefix (None / Chinese_T5)

Output:

• response: Model response text

Reset Chat Session

Reset chat session history.

Inputs:

• session_id: Session ID to reset
• trigger: (Optional) Connect any input to trigger reset

Outputs:

• success: Boolean indicating success
• message: Status message

Clear Model Cache

Clear all cached LLM models to free memory.

Inputs:

• trigger: (Optional) Connect any input to trigger cache clear

Outputs:

• success: Boolean indicating success
• message: Status message

List Chat Sessions

List all active chat sessions.

Output:

• sessions: List of active sessions with message counts

Installation

1. Place the ComfyUI-LLM-GGUF folder in ComfyUI/custom_nodes/
2. Place GGUF model files in ComfyUI/models/LLM/
3. (Optional) Install llama-cpp-python for native inference:

   pip install llama-cpp-python
   

4. Restart ComfyUI

Model Folder

GGUF models should be placed in:

ComfyUI/models/LLM/

Requirements

• ComfyUI
• llama-cpp-python (optional, for native binding)
• llama-cli.exe (for subprocess fallback if binding not available)

Author

eddy