Gemma 2 IT Dialogue

Gemma 2 IT Dialogue is the plain conversational sibling of the pack's Gemma2PromptNode. Same local gemma-2-2b-it model, same transformers loading code, but no prompt recipes and no positive/negative split - you get a prompt box and a text reply. It's the pack's "local ChatGPT" node, and it's honestly the most barebones thing in the collection: the model, a chat template, and one STRING out.

The mechanism wraps your prompt in Gemma's <start_of_turn>user ... <end_of_turn> chat template and generates with a fixed temperature of 0.7, your top_p (default 0.95), and max_new_tokens (default 100, up to 2000). device (cuda/cpu) and precision (float32/float16) are the same tuning knobs as the prompt node. There's no conversation history - each run is a fresh single-turn call, which keeps it simple but means it can't hold a multi-turn dialogue despite the "Dialogue" name. Output is a single STRING.

Installing it - the model is the install

cd ComfyUI/custom_nodes
git clone https://github.com/yiwangsimple/ComfyUI_GroqChat

Then the real work - the README tells you to manually download the model:

mkdir -p ComfyUI/models/LLavacheckpoints/gemma-2-2b-it
# download https://huggingface.co/google/gemma-2-2b-it/tree/main into that folder

Gemma 2 is gated on HuggingFace, so accept the terms and authenticate (huggingface-cli login) before the download works. transformers and torch come from the pack's requirements.txt. Restart ComfyUI after placing the model.

Gotchas

If the folder is empty the node raises "Model not found... Please manually download the model" - check the models/LLavacheckpoints/gemma-2-2b-it path specifically, and ignore the gemma-2b-it URL in the error message; it's a copy-paste inconsistency. First run is slow (model load, possibly float32 on CPU). And remember the no-history thing: if you want a back-and-forth, you'd need to chain nodes or use the API-based chat nodes instead. This node is really just "is a 2B Gemma running locally - yes." It works, but between the gated download and the single-turn limitation, the API chat nodes in this pack are the easier path for most workflows. Where this one wins is full privacy and zero ongoing cost.