Call LLM Basic

Load LLM Model Basic gives you an LLM object. This is the node that actually makes it say something. Call LLM Basic takes that object plus a prompt, runs it through the model, and hands you back a plain STRING - the closest thing ComfyUI has to "text came out of the machine."

The mechanics are deliberately thin. The pack calls it "a simplified call" of llama-cpp-python's Llama.__call__ method, and it means it: you get three knobs and a text output, no streaming, no sampling menus, nothing clever. That's the appeal. If your first LLM-in-ComfyUI workflow should be boring and reliable, this is the node to make it boring.

The inputs that matter

• LLM - feed it the output of a Load LLM Model node. This is the pack's custom type, so it only accepts wires from this pack's loaders.
• prompt - a multiline text box. This is your whole conversation with the model; there's no chat history, no system-prompt field, no multi-turn memory here. Put the entire instruction in one string, like you're writing a one-shot prompt for a fresh model every time.
• max_response_tokens - the cap on how many tokens the model is allowed to write. Default 0 effectively means "let the model decide," and the practical result is often a shorter reply than you wanted - if your output keeps getting truncated or ends suspiciously early, raise this and re-run.
• temperature - the randomness dial, default 0.8. Near 1 it gets creative and loose, near 0 it gets deterministic and repetitive. For anything that will feed into an image prompt, 0.5–0.8 is a comfortable band.
• seed - -1 means random every run; set a number to make the output reproducible.

One output, named STRING, and it's the actual generated text. That's it.

The catch: you can't see the string

Here's the gotcha that trips everyone on day one: this node's output is text, but ComfyUI's default UI has no widget that renders a raw string. Run the graph and the text is silently computed and passed along - invisible. The author's own recommendation is to pair it with ShowText from pythongosssss/ComfyUI-Custom-Scripts, which displays the value right on the node. Install that pack (one Manager search away) and wire this node's STRING into ShowText's input. From then on, "run" means "see the reply."

Install and get going

cd ComfyUI/custom_nodes
git clone https://github.com/daniel-lewis-ab/ComfyUI-Llama
# manual installs also need: pip install llama-cpp-python

Or use ComfyUI Manager (search "ComfyUI-Llama"). GGUF models go in ComfyUI/custom_nodes/ComfyUI-Llama/models, then restart and Ctrl+F5 to hard-refresh the browser. Nodes appear under the LLM menu.

If the run just fails silently, check the console: the pack's error handling and logging are described by its own author as "not mature," which is a polite way of saying a bad prompt or a context overflow can fail without a friendly message in the UI. Your first move when a Call LLM node gives you nothing is the terminal window behind ComfyUI, then the model's context size in the loader.

Where does this fit in the wider graph? The honest answer is: wherever you want it to. This pack exists because the author wanted text generation and image generation in one interface - use Call LLM Basic to generate an image prompt, a caption, a filename, or a config string, and route the STRING anywhere another node expects text. Just remember you're doing one-shot text generation, not building a chatbot.