LLM Load Model

If you're on a single GPU and your image pipeline is already fighting for VRAM, you probably don't want a 7B model cold-starting mid-run and kicking your checkpoint out of memory. LLM Load Model is the "make sure it's resident before I use it" node: you wire it a client and a model, it tells the server to load that model into memory, and it hands the model name forward so Chat Completion uses exactly that one.

Why it exists

Local LLM servers are lazy by default - Ollama loads a model on first request and unloads it when it's idle or another model shows up. That's fine for casual chatting, but in a ComfyUI queue it means the model might be loading exactly when your workflow finally reaches the chat step, eating the timeout or stalling the whole run. Loading explicitly up front makes the timing predictable.

How it works

The node needs two things: a client (LLM_CLIENT socket from Ollama Client or LlamaCpp Client) and a model. The model input is a live dropdown of the models on your connected server, so you usually don't even type - you pick. When the node runs it calls the server's load, waits for it to finish, and returns model_name (the model you chose) as its only output.

That output is the smart part. Wire model_name into Chat Completion's model input and you've enforced ordering by data dependency: ComfyUI can't run the chat before the load, because the chat needs a value the load produces. No relying on execution order luck.

Where it fits

• Ollama Client → LLM Load Model → Chat Completion → LLM Unload Model

That's the load → chat → unload chain comfydv's README describes. The load pins the model, the chat consumes it, and the unload (whose model input gets fed by the chat's model_name) evicts it afterward so the rest of your image pipeline gets its VRAM back.

Install and gotchas

Install the pack via ComfyUI Manager (search "comfydv") or:

cd ComfyUI/custom_nodes
git clone https://github.com/darth-veitcher/comfydv.git

then restart and bring your own backend (ollama serve / llama-server).

Two things to know:

• The dropdown only populates at server start. Start your model server before ComfyUI, or use the ⟳ Refresh button after pulling a new model. This node deliberately bypasses ComfyUI's frozen-at-startup validation so a freshly pulled model still works - but you do have to refresh to see it.
• This node is only worth adding when VRAM is actually tight. On a machine with lots of headroom, the lazy-loading server does the right thing on its own. The whole load/unload pair is memory-constraint tooling - use it when you feel it, not because it's on the menu.