SGLDiffusion Server Model

It doesn't load a model. The server already did that.

SGLDiffusion Server Model sounds like the node that loads the model. It doesn't. That's the first thing to unlearn about this pack's Server Mode. The model lives on the SGLang Diffusion server - a separate process, possibly on a totally different machine - and this node is just the dial tone: it creates the HTTP connection that the rest of the server-mode nodes speak over.

The sgld_client output it produces is the pack's custom connection type (SGLD_CLIENT), and every other server-mode node - Generate Image, Generate Video, Server Set LoRA, Server Unset LoRA - takes it as its first input. Think of it as the "server wire" that threads through your whole graph. You need exactly one of these, wired to everything.

The two fields you set

• base_url - defaults to http://localhost:3000/v1. That's the real SGLang Diffusion server address and port with /v1 appended; if your server runs elsewhere, put its URL here. The node is careful about the /v1 suffix - add it or not, it normalizes either way.
• api_key - defaults to the obviously fake sk-proj-1234567890. For a local server it's a formality; any string passes. If you're talking to a hosted or secured endpoint, this is where the real key goes.

The second output you'll forget

Besides the client, the node also emits model_info, a plain string. On connect it pings the server's /models endpoint and formats what comes back - model path, task type (T2I, I2I, T2V...), pipeline name, GPU count, DiT and VAE precision. It's not glamorous, but it's your first diagnostic when nothing works: wire it into a text-display node and you'll see exactly which model the server thinks it's serving. If model_info reads like an error, that's your answer before you touch anything else.

Install

Shared with the rest of the pack, and note the pack ships no requirements.txt - SGLang is a separate install:

cd ComfyUI/custom_nodes
git clone https://github.com/endman100/ComfyUI_SGLDiffusion_Fix
pip install sglang[diffusion]   # the actual heavy dependency

Restart ComfyUI after cloning. ComfyUI Manager can also handle the clone (search "SGLDiffusion") but it won't install SGLang for you - that part is manual.

Common issues

The failure mode is boring and immediate: server not running or not reachable → the model_info output shows a connection error. Check the server's own logs first, then that the port matches. One thing that trips people up: this node only makes the connection; it performs no validation that the model you plan to prompt is the one you wanted until you run a generation. And if you change the model on the server, the old client still works - just run this node again to refresh the info string. It's a plumbing node, and like most plumbing, it's invisible when it works and obvious when it doesn't.