Unsloth Desktop Nodes

comfyui-dever-unsloth

Ollama-style ComfyUI nodes for an Unsloth Desktop (llama.cpp) backend. Same 3-node simplicity as comfyui-ollama — Connectivity, Options, Generate — but talks to the OpenAI-compatible HTTP API Unsloth Desktop exposes (default http://127.0.0.1:8888, Bearer API key).

Install

Option A — ComfyUI-Manager (recommended)

1. Open the Manager menu in ComfyUI → Install Custom Nodes.
2. Search for Unsloth Desktop Nodes (registry id dever/dever-unsloth).
3. Click Install, then restart the server / refresh the UI.

Option B — Manual (clone the repo)

cd ComfyUI/custom_nodes
git clone https://github.com/DeverStyle/comfyui-dever-unsloth.git

Then restart ComfyUI. No extra Python dependencies are needed — it uses requests, which is already in the ComfyUI environment.

Nodes

| Node | Purpose | |---|---| | Unsloth Connectivity | Server URL, API key, model dropdown (🔄 Reconnect button), keep-alive. The dropdown lists models downloaded locally in Unsloth Desktop (incomplete downloads excluded) — not the full catalog of models it can use. | | Unsloth Options | Enable-flagged sampling params (temperature, top_p, top_k, min_p, repetition_penalty, presence_penalty, seed, stop, max_tokens). Only enabled values are sent. | | Unsloth Generate | system + prompt (+ optional IMAGE for vision, think, text/json format) → result, thinking, meta (for chaining). |

Requirements

• Unsloth Desktop running with Settings → "Model auto-switch (OpenAI API)" enabled. This lets a chat request name any downloaded GGUF and Unsloth loads it in-request before responding. Without it, the node raises a clear error telling you to load the model in Unsloth Desktop or enable auto-switch.
• An API key from Unsloth Desktop Settings → API (the OpenAI API requires it).

Keep-alive (Ollama-compatible semantics)

Unsloth's OpenAI API has no per-request keep_alive, so it is emulated client-side using Unsloth's POST /v1/unload:

| keep_alive | Behavior | |---|---| | 0 | Unload the model immediately after the response. | | N (>0) | A background heartbeat keeps the model warm; it unloads N (minutes/hours) after the last generation. | | -1 | Keep the model loaded indefinitely. |

Only one model is active per Unsloth server at a time; the heartbeat is keyed per server URL and follows whichever model the workflow last generated with.

Notes

• num_ctx (context size) is a load-time setting in Unsloth, not a per-request parameter, so it is intentionally not exposed here.
• No new dependencies — uses requests, already in the ComfyUI environment.