🖥️Easy Load LLM LoRA

Easy to misread this one at a glance if you're used to ComfyUI's image-generation side - this LoRA has nothing to do with Stable Diffusion or Flux. It's a PEFT-style LoRA adapter for a language model: a small set of trained weights that nudges a local LLM toward a particular style, domain, or behavior, applied on top of the base model at load time instead of retraining the whole thing. Same underlying idea as an image LoRA - small file, big behavioral shift, stacks on top of a base model - just applied to text generation instead of pixels.

How it fits into the pipeline

This node only makes sense downstream of one of this pack's local model loaders (not an API loader - you can't attach a fine-tune adapter to a closed API model you don't have weights for). model takes the model object coming out of a local loader, lora_path is a dropdown you pick your LoRA adapter from, and is_enable is the usual bypass toggle for testing with and without the adapter applied. The single output, model, is the same model object with the LoRA merged in, ready to feed into this pack's Local LLM calling node exactly like an un-adapted model would.

Worth noting: this node is flagged as an output node in the underlying schema, which in ComfyUI terms means it's set up to execute even if nothing downstream is wired to consume its output - the kind of thing you see on nodes whose real job includes a side effect (logging what got loaded, printing a status) alongside passing a value forward, rather than being a pure pass-through step you could silently skip if unused.

lora_path's dropdown is populated dynamically, the same pattern this pack uses for its local model loaders - it scans wherever your LoRA adapters live and lists what it finds, rather than shipping any built-in choices (the schema ships with zero default options). If you haven't dropped an LLM LoRA file anywhere yet, this dropdown will be empty, same as the model-picker dropdowns elsewhere in the pack when their folder hasn't been populated.

Installing it

This node ships with the pack, no separate install:

• ComfyUI Manager: search "comfyui_LLM_party", install, restart.
• Manual: cd ComfyUI/custom_nodes && git clone https://github.com/heshengtao/comfyui_LLM_party, then pip install -r requirements.txt from inside the pack folder using ComfyUI's own Python (portable builds: python_embeded\python.exe -m pip install -r requirements.txt), then restart.

Common issues & troubleshooting

lora_path is empty. That's the node correctly reporting there's nothing to pick from, not a bug - you need an actual LLM LoRA adapter file in the location this pack expects before anything shows up. It won't be in the same place as an image-generation LoRA even if both live under a folder ComfyUI calls "loras" somewhere - this is a separate model type entirely and the wrong file in the wrong place just won't appear.

Model loads fine but output doesn't seem to reflect the LoRA at all. Double check is_enable is actually on, and that the model input is genuinely coming from a local loader - wiring an API-loaded model into this node isn't a supported combination, since there are no local weights on your machine to merge an adapter into in that case.

Where to actually get an LLM LoRA. This is a bring-your-own-weights node - the pack doesn't ship any adapters, and the README's model-download links (Quark and Baidu cloud) are for base models, not LoRAs. You'd be training one yourself (with a tool like PEFT/LoRA fine-tuning scripts against your chosen base model) or sourcing one from HuggingFace for the specific base model you're running locally.