LLM Options: Textgen

LLM Options: Textgen is the most complete of the pack's three options nodes, because oobabooga Textgen exposes the whole llama.cpp-family sampler menu and this node exposes most of it back to you. It's for the same toggle-per-parameter pattern as the others: each parameter has an enable_* Boolean in front of it, everything starts OFF, and only what you enable is sent to the backend. Output is an LLM_OPTIONS object for LLM Generate (Advanced)'s options input. Disconnect it and Textgen uses whatever its own settings say.

What's on the menu

The familiar core plus a few Textgen favorites you won't find on the other two nodes:

• temperature (default 0.7), top_p (1.0), top_k (40), seed (0) - the standard quartet.
• min_p (default 0) - a newer, increasingly popular filter: drop any token whose probability is below min_p relative to the top token. It's the modern replacement for a lot of top_k fiddling, and Textgen has supported it for ages.
• repeat_penalty (1.1), presence_penalty / frequency_penalty (0) - the repetition controls.
• typical_p (1.0) - typical sampling, which trims tokens based on how "surprising" they are relative to the expected distribution. A Textgen staple.
• tfs (1.0) - tail-free sampling, the entropy-based cutoff that was all the rage in the early local-LLM days and still lives in Textgen's UI. Values below 1.0 trim more aggressively.
• max_tokens (default 512) and stop_string (mapped to the stop API parameter).

That's twelve toggles total, which is a lot of knobs - and the honest guidance is that you don't need most of them. The ones people actually reach for: temperature for general randomness, min_p as the single "crank this to clean up garbage output" control, and repeat_penalty when a model starts looping. The rest are the fun-zone of the sampling hobby, not the load-bearing defaults. Start with everything off, let Textgen's own settings do their thing, and only enable what a specific model misbehaves on.

Why it exists

Textgen is the pack's first-class citizen (the README's product direction says "Textgen-first"), so it gets the full sampler surface while LM Studio and OpenAI get trimmed sets. If you run oobabooga, this node is how you bring per-node sampling control into the graph instead of editing the Textgen UI between runs. Combined with the dedicated LLM Provider: Textgen and its integrated VRAM handling, it's the intended "everything for Textgen" stack. Unsupported parameters are dropped silently (info-level log), same as the other options nodes.

Install

Pack standard: ComfyUI Manager search "comfyui-llm-bikeshed", or clone into custom_nodes/ and pip install -r requirements.txt. Only pyyaml and requests, already in ComfyUI core - no model downloads, no config needed here. The node only builds a dict; your Textgen URL and any --api-key / --admin-key live on the provider and in config.yaml.