vLLM Connective

VLLMConnective is the pack's second connector node, and it plays the same role as OllamaConnective but for a different audience: people who already run a vLLM server. vLLM is the high-performance inference engine for big LLMs - the one you reach for when a model is too large or too slow for Ollama, or when you want a proper serving setup with continuous batching. It exposes an OpenAI-compatible API, and this node is the handshake with it.

It's the same pattern as the Ollama connector: give it a server address, it pings the server, and it packages what it learns into a VLLM_CONNECTION object plus a text summary. Leave the model field empty and it auto-selects the first model the server reports as loaded. The node is deliberately tolerant about URLs - give it http://127.0.0.1:8000 or http://127.0.0.1:8000/v1 and it appends the /v1 path correctly either way.

How it works

On execution it does a GET <base_url>/v1/models with a short timeout. vLLM answers with the list of loaded model ids; the node validates your chosen model against that list, errors with the available options if it doesn't match, and builds a connection dataclass holding the base URL, optional auth, the selected model, and the full model list. That object rides the connection output into the generator.

The inputs that matter

• base_url - default http://127.0.0.1:8000/v1. Point it at your vLLM server; remote hosts work fine too.
• api_key - default EMPTY, meaning no auth header. If your vLLM server was started with --api-key, put the key here.
• model - optional; blank auto-picks the first loaded model.

Outputs: connection (VLLM_CONNECTION) and models_text (STRING listing the selected and available models, good for a display node).

Install

It's part of the comfyui_pilcothink_VisionSLM pack:

cd ComfyUI/custom_nodes
git clone https://github.com/gpdev-Pilcothink/comfyui_pilcothink_VisionSLM

Restart, or install via ComfyUI Manager (search comfyui_pilcothink_VisionSLM). The vLLM nodes only need requests over HTTP - the heavy transformers/faiss stack in the pack's requirements is for the bundled model nodes, not for this path.

Gotchas

The connective assumes an OpenAI-compatible vLLM server is already up and serving; it won't start one for you. If the ping fails or /v1/models returns non-200, the error message includes the URL it tried - so the usual fix is checking the port and that the server actually loaded a model. Unlike Ollama, vLLM servers are often started with an API key for good reason (they're frequently network-exposed), so the EMPTY default is more likely to bite you here - if you get a 401, that's the first thing to check. And, same as the rest of this pack: new, no community mileage, install from the official repo.