orch.AcceleratorError: CUDA Initialization Failure in ComfyUI

The orch.AcceleratorError: Chyba CUDA error occurs when ComfyUI’s acceleration layer fails to initialize CUDA. This typically happens when the GPU driver version is incompatible with the CUDA runtime embedded in PyTorch, or when the GPU is not visible to the CUDA subsystem. The Czech “Chyba” simply means “Error,” indicating a generic CUDA initialization failure that prevents GPU acceleration.

Fix 1: Update NVIDIA GPU Drivers

Outdated drivers are the most common cause of CUDA initialization failures. Runaihome systems require driver version 535.x or newer for RTX 30/40 series GPUs.

# Check current driver version
nvidia-smi

# On Ubuntu/Debian:
sudo apt update
sudo apt upgrade nvidia-driver-535 nvidia-dkms-535

# On Arch:
sudo pacman -Syu nvidia

# Restart the system
sudo reboot

After reboot, verify with nvidia-smi and confirm the CUDA version displayed matches or exceeds 11.8.

Fix 2: Reinstall PyTorch with Correct CUDA Version

If drivers are current but the error persists, PyTorch’s CUDA bindings may be corrupted or built for a different CUDA toolkit version.

# Uninstall existing PyTorch
pip uninstall torch torchvision torchaudio -y

# Reinstall with CUDA 11.8 support (ComfyUI v0.2.x compatible)
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu118

# Alternatively, for CUDA 12.1 (newer systems):
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121

Verify installation: python -c "import torch; print(torch.cuda.is_available())" should return True.

Fix 3: Set CUDA_VISIBLE_DEVICES Environment Variable

On systems with multiple GPUs or when running in containerized environments, CUDA may fail to detect the correct device.

# List available