Local LLM Settings (GGUF)

Every LLM node in the Kinburg pack - Local LLM (GGUF), the chat node, Token Counter, Context Sizer - takes its orders from one place: Local LLM Settings (GGUF). It's the config node that carries the model, the system prompt, the sampling, the context size, the reasoning control, and the output format, then emits the whole lot as a single config bundle. One of these can drive several LLM nodes at once, which means you change the model once and every consumer follows. It's the node you'll spend your time on, so it's worth knowing what's actually in it.

What you'll actually set

• model / model_path - pick the .gguf from ComfyUI/models/llm (subfolders show as folder/model.gguf), or type a path in model_path via the placeholder. Quotes from Windows "Copy as path" are stripped automatically.
• system_prompt - who the model is. This is where the job gets defined.
• max_tokens, temperature, top_p, top_k, min_p, repeat_penalty, stop - the sampling basics. min_p of ~0.05 is a good start, often paired with top_p=1.0, top_k=0. Stop strings go one per line.
• n_ctx - context size. The KV cache shrinks with it, so don't just max it out.
• n_gpu_layers - -1 = all layers on GPU, 0 = all CPU.
• flash_attn, kv_cache_type - flash attention is faster and uses less VRAM; quantizing the KV cache (q8_0/q4_0) fits a bigger context in the same card. Both auto-enable each other as needed.
• unload_comfy_models - frees image models from VRAM before the LLM runs. unload_llm_after_run - free the LLM after each run; off (default) keeps it warm for chat, turn it on in image workflows.

Reasoning control - the modern stuff

Modern reasoning models are switched inside the chat template, not by anything you say in the prompt. enable_thinking (model default / on / off) and reasoning_effort are handed to the model's own Jinja template as variables, so the model can't ignore them. Big caveat: model default is not off - Qwen3.5/3.8 take an undefined variable as thinking ON, Gemma-4 as OFF. The older thinking_directive (appends /no_think to the prompt) only works on original Qwen3; newer models ignore it. strip_think keeps reasoning out of the text output (it still goes to thoughts), and answer_marker handles models that reason without <think> tags.

Structured output & advanced

output_format - free text, valid JSON, a custom GBNF grammar, or ideogram4_json (a built-in prompt grammar). Grammar runs skip the live progress bar. extra_load_args passes Python-binding args to llama-cpp-python's Llama() loader (n_threads=8, main_gpu=0, tensor_split=[1,1]), one key=value per line - note these are binding args, not llama.cpp CLI flags, so --spec-type won't work here. chat_template_path overrides a broken embedded chat template; leave it empty for well-packaged models. Optional connect-only inputs: context (reference material appended to the system prompt) and vision (from a Vision Settings node). Output is just config.

Install

Pack install via ComfyUI Manager (search "Kinburg-Nodes") or git clone. Needs llama-cpp-python - Manager runs install.py automatically, or run <ComfyUI>/.venv/Scripts/python.exe <ComfyUI>/custom_nodes/Kinburg-Nodes/install.py by hand. The wheel's CUDA major must match your torch's, or you get the cudart64_12.dll load failure. One config node to change, several LLM nodes that follow - that's the whole design, and it's a good one.