ComfyUI-DaSiWa-Nodes

DaSiWa LLM Model Selector: choose a local transformers LLM/VLM folder from ComfyUI/models/llm and configure caching, dtype, device, and unload behavior.

Inference backend. Transformers is supported in this version.

Cached is faster. Unload after run frees RAM/VRAM after every output.

Optional absolute path, or relative path under ComfyUI/models/llm. Overrides model when set.

Download hf_repo_id into ComfyUI/models/llm if it is not already present.

Model dtype. Auto follows the model config when possible.

Optional Hugging Face repo id, for example Qwen/Qwen2.5-VL-7B-Instruct. Overrides model when set.

HF branch, tag, or commit. Used when downloading or locating a repo copy.

Model folder under ComfyUI/models/llm. Use a full Hugging Face-style folder with config/tokenizer files.

Optional bitsandbytes quantization. Requires bitsandbytes installed.

Use vision when analyzing connected images/frame batches.

Allow custom model code from the model folder.

ComfyUI Node: DaSiWa LLM Model Selector

Category

Inputs

Outputs

Extension: ComfyUI-DaSiWa-Nodes

More nodes in ComfyUI-DaSiWa-Nodes

Run ComfyUI workflows without the setup