Nova ACE Setup Check 🩺

Somewhere between "I have a folder of songs" and "why is this OOMing" is a whole class of problems you can find in one second instead of one hour. Wrong checkpoint layout, a variant folder that doesn't exist, missing training packages, a torchcodec built for the wrong backend.

Nova ACE Setup Check 🩺 answers exactly one question: would a preprocess or training run start on this machine? It reports. It does not install, download, or spawn anything - the author is explicit about that, and it's the right call for a node that might sit in a saved workflow for months.

What it inspects

Give it a checkpoint_dir and a variant and it checks the HuggingFace-layout tree for that variant: the folder, config.json, the remote-code Python files, model.safetensors, plus vae/, Qwen3-Embedding-0.6B/ and silence_latent.pt. It checks torchcodec, the ACE-Step training packages, the optional ones (bitsandbytes for adamw8bit, prodigyopt for prodigy, tensorboard), which compute backend torch was built for, and whether the GPU is actually visible.

Two of its checks call the same functions Preprocess and the Trainer call, which is the bit that makes it trustworthy: a green result here means those nodes won't stop on the same grounds. Anything missing is collected into a single install command built for whatever your environment actually has - pip or uv. Packages ComfyUI itself ships are never suggested; if it says torch is missing, that's a broken install or the wrong interpreter, and pip install --no-deps torch would make things considerably worse.

Inputs and outputs

• checkpoint_dir - the checkpoint tree root. Leave it empty and it checks everything except the tree, which is a genuinely useful mode when you're still downloading.
• variant - xl_sft by default; xl_turbo, xl_base, sft, turbo, base are the rest. Match what you intend to train.
• acestep_repo_path (optional) - checked for importability. Note the tooltip: the trainer needs it set on its own node too, because a subprocess doesn't inherit sys.path. Setting it here does not set it there.

Outputs: ready (a boolean - true only when nothing would stop a run; advisories like "no tensorboard" don't clear it) and console, the full report. Wire the console into Nova Console and read it.

Install

ComfyUI Manager → Nova Audio Player. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/NovaFemme/ComfyUI-NovaAudioPlayer.git

Restart, and it's under ▶️ Nova Audio → 🎓 LoRA Training. Nothing else is needed to run the check itself - that's the point of it. It needs no ACE-Step clone, no models, no extra packages, so it'll happily tell you what's missing on a machine where nothing is set up yet.

What each failure actually means

NOT READY on the checkpoint tree is usually a variant mismatch, not a broken install. Check that variant here matches what you set on Preprocess and the Trainer before you re-download 9.3 GB. This is the most common false alarm, and it's also the exact mistake that ruins a run quietly later.

Missing packages - run the command it prints. If you installed through Manager, the pack's install.py already ran nova_ace_setup.py to do the package half; only the models are left. Clone-by-hand users get nothing installed, so this is where you find that out.

ROCm, Intel, or CPU-only PyTorch, torchcodec fails - that's expected and fixable. The PyPI wheel is a CUDA build and can't load against a non-CUDA torch, which breaks audio decoding everywhere, not just here:

<venv>/bin/python -m pip install --no-deps --reinstall \
    --index-url https://download.pytorch.org/whl/cpu torchcodec

12GB VRAM or less - the check has no opinion, but the reference run peaked at 11.3 GiB for six tracks at 300s max. VRAM scales with your longest track, not your track count, so a long track is what pushes you over.

Run it once before every training session where you've changed something - a new variant, a new venv, a moved folder. It's a second of your time against an afternoon of GPU.