Tara Preset LLM Config Node

TaraPresetLLMConfig is the comfort-food version of TaraLLMConfig. Where the plain config node makes you type a base URL and a model name, this one gives you a dropdown of the pack's model list - the same openai/gpt-3.5-turbo, groq/mixtral-8x7b-32768, groq/llama3-8b-8192 menu the deprecated nodes used - and resolves the endpoint for you. It still outputs the modern TARA_LLM_CONFIG object, so it plugs into TaraPrompterAdvanced and TaraAdvancedComposition exactly like the manual config does. If you just want to get Groq working in thirty seconds without thinking about base URLs, this is your node.

How it works

Pick a model from the dropdown and the node splits it on the /: the part before is the provider, the part after is the model name. It then hardcodes the right base URL for the provider - Groq gets https://api.groq.com/openai/v1, OpenAI gets https://api.openai.com/v1 - bundles your sampling settings, and emits llm_config.

The inputs that matter:

• llm_models - the dropdown. Provider + model in one string.
• use_loader (default on) - when on, the node calls TaraApiKeyLoader internally to fetch the saved key (honoring loader_temporary). When off, it reads the optional api_key input instead.
• temperature (0.4), seed (42), max_tokens (1000), top_p (1), frequency_penalty / presence_penalty (0), timeout (60) - the same sampling knobs as the manual config node.

So the two ways to feed a key: let it pull from the Saver's API_KEYS.json (secure-ish, but requires the Saver to have run), or turn use_loader off and wire a key in directly. The latter is the README's recommended pattern for shared/hosted boxes where you don't want keys saved to disk at all.

The two ways this node bites you

First, together/coming-soon is not a real option. It's in the dropdown because it's in the shared model list, but the preset's provider resolver only knows Groq and OpenAI - pick Together and the node throws a ValueError. It's a placeholder for a feature that never shipped. Second, if use_loader is on and you haven't run TaraApiKeySaver yet, you get FileNotFoundError: API key not found. Both fail loudly, which is honestly better than silently producing a config with an empty key. Also keep the temporary flags consistent between Saver and loader (loader_temporary here) or the two look in different directories and you'll swear the key vanished.

Install and placement

cd ComfyUI/custom_nodes
git clone https://github.com/ronniebasak/ComfyUI-Tara-LLM-Integration

Restart, or Manager → Install via Git URL. Deps: orjson and openai, no models. Wire it like the manual config - llm_config into TaraPrompterAdvanced or TaraAdvancedComposition - and you're done. The one real limitation is the frozen model list: it's early-2024 vintage, so when you outgrow llama3-8b, the freeform llm_model field on plain TaraLLMConfig is waiting for you.