ACE-Step Clear VRAM

Every node in this pack loads a big diffusion model, and the base trio (Extract/Lego/Complete) plus the omni models are hungry enough that running several in one session is a reliable way to hit out-of-memory errors. Clear VRAM is the pack's answer: a node that sits in your graph and, when it runs, unloads the ACE-Step models and forces a full GPU memory sweep. Its entire job is "make the next thing fit."

It's honest about being a plumbing node. It has zero required inputs - the only input, any_input, exists purely as a pass-through so you can chain it into a workflow without breaking your data flow, and it returns that same value as pass_through. It's marked as an output node, so it runs even when nothing else is connected to it, and it prints its work to the console: memory before/after, which ACE-Step handlers it cleared (dit.model, dit.vae, llm.model), and total freed.

How it works

Under the hood it does what you'd do by hand, but more thoroughly: it tears down the pack's cached ACE-Step DiT and language-model handlers, then runs three rounds of garbage collection plus torch.cuda.empty_cache(), resets peak memory stats, and finally calls into ComfyUI's own model_management.free_memory() and soft_empty_cache() so the rest of your session benefits too. The three-pass cache clearing matters - one empty_cache() often isn't enough after a big diffusion render.

The inputs that matter

There are none, really. That's the point. If you want it to not disrupt a workflow, wire your audio (or anything) into any_input and it flows out the other side untouched; otherwise just drop the node in and let it run.

Install

Part of the standard pack install - nothing separate:

git clone https://github.com/ACE-Step/ACE-Step-1.5.git
cd ACE-Step-1.5 && pip install -e .
cd ComfyUI/custom_nodes
git clone https://github.com/kana112233/ComfyUI-kaola-ace-step.git
cd ComfyUI-kaola-ace-step && pip install -r requirements.txt
huggingface-cli download ACE-Step/Ace-Step1.5 --local-dir ComfyUI/models/Ace-Step1.5

Restart ComfyUI after installing.

Where people get burned

The main trap is expecting it to be a magic fix for all OOMs - it clears what this pack loaded, and it genuinely helps with the pack's own model stacking, but it can't conjure VRAM you don't have. The realistic play is to think of it as a workflow-level release valve: run a heavy ACE-Step generation, then chain Clear VRAM before the next stage so the following node starts with a clean board. It pairs naturally after Extract or the base-model nodes, which is exactly how the pack's own example workflow (extract-clean-VARM.json) uses it. If you're still OOMing after it runs, the answer is a smaller model or lower batch size, not more cleanup.