Simple Qwen Unload

Here's the tension at the heart of this pack: the author built the whole thing so that after the LLM runs, it leaves nothing in VRAM - and then added keep_vram and save1–save3 modes that deliberately keep the model loaded for speed. Those two goals need a manual off-switch, and that off-switch is Simple Qwen Unload.

The node does two things at once. First, it force-unloads a cached Qwen model from VRAM - targeting either all caches or a specific one (keep_vram, save1, save2, save3). Second, it's a pass-through: the input port flows through to output unchanged, so you can drop it into a data path and it both unloads the model and sequences execution at the right moment.

Inputs and outputs

• target - which cache to unload: all, or the specific keep_vram / save1 / save2 / save3 cache. Default is all. If you've got several small models parked in the save slots, all is the clean sweep; if you only want to drop one, pick it by name.
• input - anything, passed through untouched.
• output - the same value, after the unload happens.

When you need it

Three situations, from the README and the pack's own design:

1. You used keep_vram or save modes. The model stays resident by design. When the LLM stage is done and a video model or big sampler needs every byte, this node is how you get the memory back. The save modes in particular have no auto-cleanup - the README is explicit that the only way to unload them is this node. Forget it and you'll be wondering why your next generation is crawling or OOMing.
2. Batch pipelines. Run a series of LLM inferences, then unload once, before the heavy part of the workflow.
3. Reset before a different config. Swapping to a model with different settings is cleaner if you unload the old cache first.

The execution-ordering bonus is real: because the node passes its input through, you can wire it between the LLM's output and whatever runs next, and you've guaranteed the unload happens before the next heavy node - not after it's already fighting for VRAM.

Install

It ships in KLL535/ComfyUI_Simple_Qwen3-VL-gguf, same as the main Qwen node:

cd ComfyUI/custom_nodes
git clone https://github.com/KLL535/ComfyUI_Simple_Qwen3-VL-gguf

or ComfyUI Manager (search ComfyUI_Simple_Qwen3-VL-gguf), restart, F5.

A fair warning on the workflow side: don't get clever and unload aggressively in subprocess/direct_clean mode - those modes already free the model after every run, so an unload node there is redundant at best and can force an avoidable reload at worst. The node earns its keep specifically when the model is supposed to stay loaded until you say otherwise.