🖥️Easy Local LLM Loader

Of comfyui_LLM_party's several local-model loaders, this is the one aimed at people who don't want to hand-type a filesystem path or a HuggingFace repo ID and get it wrong. The GGUF Loader and VLM Loader both take a free-text model_path / model_name_or_path field where a typo just gives you a load error with no hints. This one instead scans a fixed folder - custom_nodes\comfyui_LLM_party\model\LLM - and turns whatever's in there into an actual dropdown you pick from. That's the tradeoff: less flexible about where your model files live, but you can't fat-finger the path.

The catch, obviously, is that you have to put a model there yourself first - this node doesn't download anything, it just discovers what's already sitting in that folder. Grab a compatible model from HuggingFace (the README's own tested list includes things like Qwen/Qwen2-7B-Instruct and meta-llama/Llama-2-7b-chat-hf), drop the model folder into comfyui_LLM_party/model/LLM, restart or refresh ComfyUI, and it'll show up in model_name_or_path's choices. If the dropdown comes up empty, that's not a bug - it's telling you the folder is empty, and choices: [] in the underlying schema confirms there's genuinely nothing hardcoded to fall back on.

Beyond the path handling, this is a straightforward transformers loader: device (auto/cuda/cpu/mps) picks where it runs, dtype (auto/float32/float16/bfloat16/int8/int4) picks the precision - drop to int8 or int4 if you're tight on VRAM, same tradeoff as any quantized transformers load. is_locked is the memory-management flag, and its own tooltip spells out the intended pattern more clearly than most: locked, the model stays loaded and won't reload on every run; unlocked, you're expected to pair it with the pack's Clear Model node to actually release GPU memory before loading something else. Two outputs come out - model and tokenizer - both feeding into the pack's Local LLM node with model_type set to LLM.

Installing it means installing the pack: search "comfyui_LLM_party" in ComfyUI Manager, or git clone https://github.com/heshengtao/comfyui_LLM_party into custom_nodes and restart, then pip install -r requirements.txt from the pack folder using ComfyUI's own Python - on a portable/embedded build that's specifically python_embeded\python.exe -m pip install -r requirements.txt, since your system pip won't touch ComfyUI's own environment.

This pack shows up organically when people ask how to run local LLMs inside a ComfyUI graph - it's a real, actively used option alongside simpler single-purpose alternatives like comfyui-ollama, and this loader specifically is the friendliest on-ramp of the bunch precisely because it removes the path-typing step.

The one thing that catches people out: the folder path is fixed and case-sensitive on some systems - model\LLM, not models\LLM or model\llm. If you've correctly downloaded a model but it's just not appearing in the dropdown, double-check you put it in the exact folder this node is scanning rather than wherever felt natural, and that the folder contains the actual model files (config, weights, tokenizer files) rather than a nested subfolder one level too deep.