Safetensors -> GGUF (llama.cpp)

GGUF is how a chat LLM shares a consumer card with the diffusion model - quantized weights in one file, from Q8 down to Q4_K_M, the accepted compromise for a 12GB card. But not every model you want to run ships as GGUF; plenty of HuggingFace checkpoints come as .safetensors, and until now converting them meant a terminal session and a llama.cpp checkout. Safetensors → GGUF (llama.cpp) does the conversion from inside ComfyUI - and it's the pack's language-model converter, so it turns a Qwen or Llama HF model into a GGUF you can load straight into the pack's Local LLM (GGUF) nodes.

What it takes

• source - a HuggingFace repo id (Qwen/Qwen2.5-0.5B-Instruct), an https://huggingface.co/owner/name URL (downloaded for you), a local HF model folder (config.json + tokenizer + *.safetensors), or a single .safetensors file (its folder is used). Windows "Copy as path" quotes are stripped.
• output_dir / output_name - where the .gguf goes and its base name. Blank output_dir defaults to ComfyUI/models/llm; blank name derives from the model. The precision/quant tag and .gguf are added automatically.
• outtype - the precision written by convert_hf_to_gguf.py. f16 is the usual base. When you're quantizing, this is the intermediate step.
• quantize - the optional second pass with llama-quantize (Q4_K_M, …). none keeps the outtype GGUF and needs no binary at all.
• quantize_binary / quantize_binary_path - the compiled llama-quantize executable, auto-listed from ComfyUI/models/llm. Only used when quantize != none.

Optional: llama_cpp_dir (a llama.cpp checkout holding the convert script), auto_clone (git-clone llama.cpp under ComfyUI/models/llm automatically - needs git on PATH; clones source only, no compiled binary), hf_token (for gated repos), hf_download_dir, keep_intermediate (keep the intermediate outtype GGUF when quantizing), and force (re-run even if the target exists - off returns the existing file, which is a nice idempotence touch).

Outputs: gguf_path (the finished file, ready to wire into a loader), log (the conversion tail), and help.

The honest caveats

The dependencies are already in ComfyUI (huggingface_hub, gguf, torch, safetensors) and the scripts are fetched for you, so the friction is low - but the quantization pass needs an external llama-quantize binary you compile yourself. Leaving quantize = none needs nothing extra and is the easy first run. Also: this node is for language/multimodal LLMs only. It can't read diffusion weights - that's the pack's separate Diffusion Safetensors → GGUF (city96) node, which drives city96's ComfyUI-GGUF tooling (and its Q ladder is exactly the GGUF story: Q8 ≈ fp16 at half the size, Q5 the last stop before visible loss).

Install

Part of the Kinburg-Nodes pack - ComfyUI Manager (search "Kinburg-Nodes") or cd ComfyUI/custom_nodes && git clone https://github.com/Kinburg/Kinburg-Nodes, restart. No heavy dependencies beyond what ComfyUI ships. If you only ever run pre-quantized models, you don't need this node - but the moment a model you want only comes as safetensors, it beats fighting llama.cpp's scripts by hand.