DavchaLLM

A local LLM inside ComfyUI means you can rewrite prompts, generate tags, or summarize right in the graph - no API keys, no network, no middleman. DavchaLLM is the no-frills end of that: feed it a loaded model, a text prompt and a seed, and it hands back the model's reply as a string. Nothing more.

It's part of comfyui_davcha, the author's "personal QoL and experimental nodes" pack, and this one shows its age. In the current source it's actually registered under the name OldDavchaLLM - the author moved on to newer LLM nodes and left this one behind under the "Old" prefix. If you search the node menu in a fresh install and can't find "DavchaLLM", look for "OldDavchaLLM"; both names describe the same node.

How it works

Behind the scenes it's llama.cpp via the llama-cpp-python bindings. It takes the model you loaded, sets the seed for reproducibility, and calls create_chat_completion with a single user message - no system prompt. The sampling parameters are hardcoded: 512 max tokens, temperature 1.0, top_p 0.9, top_k 50, repeat_penalty 1.2. Those are sane llama.cpp defaults, but they're not tunable here, which is exactly why DavchaLLMAdvanced exists in the same pack.

Inputs: model (from the pack's LLM loader), seed (INT, 0 for "no fixed seed"), text (multiline prompt). Output: text (STRING) - the model's reply, ready to feed a text encode node or a string utility.

Installing it

# ComfyUI Manager → Install Custom Nodes → search "comfyui_davcha" → Install → Restart
# or manually:
cd ComfyUI/custom_nodes
git clone https://github.com/dchatel/comfyui_davcha
cd comfyui_davcha
pip install -r requirements.txt

This node is where the pack's install catch bites hardest. nodes.py imports from llama_cpp import Llama at the top of the module, but llama-cpp-python is not in requirements.txt (which only lists webp and rapidfuzz). If you don't have it, the entire pack fails to import and every davcha node disappears. Install it yourself: pip install llama-cpp-python (a chunky build - wheel or compile) plus opencv-python for the other missing import, then restart.

Where people get burned

You also need a model to load - a .gguf file in ComfyUI/models/llm_gguf/. The loader scans that folder at startup, so no file, no dropdown. Pick a quantization you can actually fit: Q4_K_M is the accepted compromise for a 12GB card, and you can forget running this class of loader on a machine with no VRAM to spare, since the loader pushes all layers to the GPU by default. And remember this is the crippled twin: if your task needs a system prompt or temperature control, you want DavchaLLMAdvanced instead.