Qwen Chat (Multi-turn)

Here's a node that has nothing to do with MiniMax H3 at all. It's a plain multi-turn chat with a local Qwen GGUF, tucked into a pack about video prompts - and honestly it's the pack's sleeper. Once you've got a Qwen model loaded for prompt-writing, you might as well use it to brainstorm the prompt ideas themselves. This is that: type a sentence, get an answer, and it remembers the conversation.

It's completely local and needs no API key - same stack as the rest of the pack, just pointed at a conversation instead of the H3 format.

How it works

The node keeps a session per node instance in memory - a list of system/user/assistant messages keyed by the node's unique_id. Each run appends your user_input, trims to the most recent max_turns (dropping the oldest turns once you exceed the limit), and calls create_chat_completion through ComfyUI-llama-cpp_vlm with the whole session. Then it appends the reply and hands you back the response plus the full formatted history.

Two niceties worth calling out. max_turns caps context growth so a long chat session doesn't balloon past your n_ctx - the tooltip spells it out: keep the most recent N turns, discard the oldest beyond that. And enable_thinking flips the thinking mode on the model's chat handler at runtime without reloading the checkpoint, which is a real Qwen3.5-style feature: off for speed, on when you want the better answer.

The inputs that matter

• user_input - what you're saying this turn. Empty input just returns the current history.
• system_prompt - empty by default, which falls back to a basic helpful-assistant system prompt. Set it to make this a specialized assistant (e.g. a video-idea critic) without touching the model.
• max_turns (default 10, up to 50) - how much of the conversation it remembers.
• clear_history - tick it and run once to wipe the session.
• enable_thinking - off by default for speed; on for thinking models.

The standard sampling controls (temperature, top_p, top_k, max_tokens, seed) behave exactly as you'd expect, and model / mmproj / chat_handler / n_ctx are the same shared LLM-loading block as the rest of the pack.

Two outputs: response (the latest reply, wire it into a text display) and history (the whole session, formatted with system/user/assistant markers - handy for dumping to a file or feeding a later LLM stage).

Installing it

Same shared setup as the pack's other nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/wangminxing2019/ComfyUI-MiniMaxH3-QwenPrompt.git
git clone https://github.com/JamePeng/ComfyUI-llama-cpp_vlm.git

Restart, and make sure a Qwen GGUF is in ComfyUI/models/LLM. For chat you don't need a vision model or an mmproj - any Qwen GGUF works.

Gotchas worth knowing

• History lives in RAM, tied to the node instance - the session is keyed to the node's id in the current workflow. Duplicate the node and you get a separate fresh session; load a different workflow and it's gone. Don't treat it as a chat log.
• A failed run rolls back cleanly - if inference throws, the node pops your user input off the session so a crash doesn't poison the conversation. Nice touch, but it means the failure message alone won't tell you what broke.
• It shares the loaded model - since it uses the same storage as the prompt nodes, switching models here unloads the one the prompt node was using. Fine, just be aware changing model or n_ctx on any of these nodes kicks the others out of VRAM.