For those of you running both ComfyUI and local LLMs, how are you managing GPU resources?
retained text
I’m curious how people here switch between ComfyUI workflows and local LLM inference, especially on a single GPU. Are you manually unloading models, using something like LM Studio or llama-swap, relying on ComfyUI’s model management, or have you built something to handle switching automatically? I’m particularly interested in how you handle VRAM contention when you want to move between image generation and LLM workloads without constantly waiting for models to reload. What’s your current setup, and what works well for you? Video attached is my project [Berth](https://github.com/T-Crypt/berth) but I’m here curious on how everyone else solved this problem?
document fetched 2026-11-03 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xln18y/for_those_of_you_running_both_comfyui_and_local