---
title: "Generate Images and Videos with ComfyUI — Troubleshooting"
canonical: "https://build.nvidia.com/spark/comfyui/troubleshooting.md"
---

# Common issues

| Symptom | Cause | Fix |
|---------|-------|-----|
| PyTorch CUDA not available (Image Gen Quick Start) | Incorrect CUDA wheels or missing drivers | Verify `nvidia-smi` and `nvcc --version`, reinstall PyTorch with the CUDA index URL from that tab |
| "permission denied" when running docker (Video Gen Workflow) | User not in docker group | Run `sudo usermod -aG docker $USER && newgrp docker` |
| Container fails to start with GPU error (Video Gen Workflow) | NVIDIA Container Toolkit not configured | Run `nvidia-ctk runtime configure --runtime=docker` and restart Docker |
| ComfyUI web UI not accessible | Firewall, wrong IP, or server not ready | Check host terminal output (Image Gen Quick Start) or `docker logs comfyui` (Video Gen Workflow); open port 8188; use `http://<HARDWARE_IP>:8188` |
| Model download fails | Network, disk space, or auth | Check connectivity, free disk, and (Video Gen Workflow) a valid `HF_TOKEN` plus accepted model licenses |
| HuggingFace download fails with 401 (Video Gen Workflow) | Invalid or missing HF token | Export a valid token from [huggingface.co/settings/tokens](https://huggingface.co/settings/tokens) |
| "Model file not found" when running a workflow (Video Gen Workflow) | Weights missing or volume mount wrong | Confirm files under `./models/` and the `-v` mounts on `docker run` |
| Out of GPU / unified memory during generation | Resolution, frames, or model too large | Use a smaller model, lower resolution, or fewer frames; see Video Gen Workflow tier table |
| Workflow loads but nodes show red "missing" (Video Gen Workflow) | Custom node not installed | Use ComfyUI-Manager → Install Missing Custom Nodes, or rebuild the image |
| Web UI: **"Error: the workflow does not contain any nodes"** on **Load** | File is API format, not a UI workflow | Load `assets/workflows/<name>.json`; use `assets/workflow_api/<name>.api.json` only with `POST /prompt` |
| `device >= 0 && device < num_gpus INTERNAL ASSERT FAILED` (Video Gen Workflow) | `--gpus all` on a multi-GPU system | Use `--gpus '"device=N"'` for the intended GPU index from `nvidia-smi` |
| NGC image pull requires authentication (Video Gen Workflow) | NGC registry login required | Run `docker login nvcr.io` with your NGC API key |
| Very slow generation, low GPU utilization | Process not on GPU | Image Gen Quick Start: confirm CUDA in the venv. Video Gen Workflow: `docker exec comfyui nvidia-smi` |
| Memory pressure on unified-memory hardware even within capacity | Buffer cache not released to the GPU | Flush the buffer cache (see note below) |
| Container exits with `ModuleNotFoundError: torchaudio` or torchaudio ABI / undefined-symbol errors (Video Gen Workflow) | Image missing the playbook torchaudio stub, or a real torchaudio wheel was layered on NGC PyTorch | Rebuild from the shipped `assets/Dockerfile`. Do **not** `pip install torchaudio` inside the container |
| Custom-node build or DWPose warns about `onnxruntime` / no GPU providers on ARM64 (Video Gen Workflow) | `onnxruntime-gpu` has no aarch64 wheel on PyPI | The shipped Dockerfile substitutes CPU `onnxruntime`. Preprocessors run on CPU; treat as informational unless nodes fail to load |
| Startup shows `aimdo` hook failures or `urllib3` / `charset_normalizer` version warnings (Video Gen Workflow) | NGC base-image diagnostics / dependency skew | Benign if the ready line `To see the GUI go to: ...` appears; ComfyUI still works |

> [!NOTE]
> Some hardware platforms such as DGX Spark use Unified Memory Architecture (UMA), which shares memory dynamically between GPU and CPU. If you hit memory errors even when within total capacity, flush the buffer cache:
>
> ```bash
> sudo sh -c 'sync; echo 3 > /proc/sys/vm/drop_caches'
> ```

> [!NOTE]
> Video Gen Workflow logs: `docker logs -f comfyui`. Most missing-model and node errors appear there with clear messages.