⚙️对话增强设置@炮老师的小课堂

This node looks like nothing much - a pile of sliders and one output - but it's the personality of the whole local chat workbench in this pack. DapaoChatSettings is where you set the system prompt, the sampling, and the image-size cap, and it hands all of it to DapaoMultiTurnChatV2 as one tidy bundle. Think of it as the control panel pulled out of the chat node so you can change the conversation's character without touching the big terminal node.

It's a pure config node. There's no model inside, no VRAM, no inference. You set values, and the single ⚙️对话设置 output (type DAPAO_CHAT_SETTINGS) plugs straight into the chat node's ⚙️对话设置 input. Everything downstream - history trimming, token caps, temperature - reads from that bundle.

The inputs that actually matter

Most of the required inputs have sane defaults you can leave alone, but two are worth a second look:

• 📝系统提示词 - the only one you'll touch every session. It's multiline, and the pack's placeholder notes something important: when you connect a Skill, the Skill owns the system rules, so keep this empty or let it defer.
• 📐图片最大边长 (default 2048) - the tooltip is the whole story: each image is Lanczos-scaled to no more than 2048px on the long edge and sent as PNG. It's your VRAM knob for vision-heavy chats. Images get bigger than that and every turn pays the price in context tokens.

Then there's the sampler cluster - 🌡️温度, 🎯top_p, 🔝top_k, 🔁重复惩罚, 📈频率惩罚, 📍存在惩罚 - plus 🔢最大历史轮数 (up to 100 rounds) and 📤最大生成token (default 1024). 🧠输出think块 toggles whether the model's reasoning stays in the reply for thinking-capable models.

Where people get burned

The classic mistake is leaving this node floating on the canvas. An unconnected settings node changes nothing - the chat node only sees it through the wire. If you edit it and the chat doesn't change, check that connection first.

The other gotcha is round count. 最大历史轮数 caps how many past turns get re-fed to the model, but remember the image trick from the pack's docs: only the images you explicitly @ this turn are processed. History stores material numbers and text analysis, not the raw media, so a long chat doesn't silently balloon your context with old pictures. That's by design and it's worth knowing before you blame the settings node for a memory problem.

Installing it

This node ships in paolaoshi/ComfyUI-llama_Dapao, installed once for all the nodes in the 🍭大炮-llama-cpp category:

cd ComfyUI/custom_nodes
git clone https://github.com/paolaoshi/ComfyUI-llama_Dapao

or search "ComfyUI-llama_Dapao" in ComfyUI Manager. Then install dependencies and restart. The pack depends on llama-cpp-python from the JamePeng fork at a pinned commit (0.3.47+, needed for Qwen3.8's MTP layers - stock PyPI's 0.3.35 fails with missing tensor 'blk.64.ssm_conv1d.weight'). CPU users install from source; NVIDIA users should overlay a matching GPU wheel from the JamePeng releases page and then restart ComfyUI. GGUF models and their mmproj files go in ComfyUI/models/LLM/.

That's it. Connect it to the chat node, wire a model from DapaoLocalModelLoader, and the whole workbench runs on your card with no API key anywhere in sight.