🦙 ComfyUI Llama-Swap Client

A native ComfyUI node for llama-swap — hot-swap any llama.cpp model without leaving your workflow.

---

✨ Features

🔄 Live model picker	Fetches /v1/models from your running server and shows a floating dropdown — click to set
🖼️ Vision support	Connect any ComfyUI IMAGE node; the first frame is base64-encoded and sent as image_url
🧠 Thinking extraction	<think> / <thinking> blocks are stripped from response and surfaced in a separate thinking output
⏏️ Auto-unload toggle	Calls /unload automatically after every generation — great for VRAM-constrained setups
📋 Running button	Shows which model is currently warm in GPU memory via a toast notification
🔴 Unload All button	Manually frees VRAM from inside ComfyUI without touching the terminal

---

🗂️ Nodes

🦙 Llama-Swap Client

The main inference node.

Input	Type	Description
server_url	STRING	llama-swap base URL (default http://localhost:8080)
model	STRING	Model name — populated via 🔄 Fetch Models
system_prompt	STRING (multiline)	System role message
prompt	STRING (multiline)	User message / question
unload_after_generate	BOOLEAN	Auto-call /unload after every run
image (optional)	IMAGE	Vision input — first frame sent as JPEG base64

Output	Description
response	Clean human-readable text — <think> blocks removed
thinking	Extracted reasoning chain (empty string if the model produced none)

---

🦙 Llama-Swap Model Selector

A standalone picker that outputs model_name as a STRING.
Useful to share the same model choice across multiple inference nodes.

---

⚡ Installation

cd ComfyUI/custom_nodes
git clone https://github.com/yourname/comfyui_llama_swap

Dependencies: requests and pillow — both already present in any standard ComfyUI environment.

Restart ComfyUI after cloning.

---

🚀 Quick Start

1. Add a 🦙 Llama-Swap Client node
2. Set server_url to your llama-swap address
3. Click 🔄 Fetch Models → select a model from the dropdown
4. Connect a Preview Text node to response
5. (Optional) Connect a second Preview Text node to thinking to debug reasoning chains
6. Hit Run 🎉

---

🧠 Thinking Output

Models like DeepSeek-R1, QwQ, Qwen3 and other reasoning models wrap their chain-of-thought in <think> tags.
This node automatically separates them:

response  →  clean answer, ready to use downstream
thinking  →  full reasoning trace for inspection / debugging

Both <think> and <thinking> variants are handled.

---

🔌 Backend Routes

Three lightweight proxy routes are registered on ComfyUI's PromptServer at startup to avoid CORS issues:

Route	Proxies to
GET /llama_swap/models	GET {url}/v1/models
GET /llama_swap/running	GET {url}/running
GET /llama_swap/unload	GET {url}/unload

---

📄 License

MIT