Why Torch Falls Back to CPU in ComfyUI (And How to Fix It)

The Symptom

You have an NVIDIA GPU, but ComfyUI generates at a crawl. Check:

python -c "import torch; print(torch.cuda.is_available())"

Prints False. Your GPU is idle.

Reason 1: PyTorch Installed Without CUDA

A plain pip install torch gives you CPU-only. You need:

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu126

Check your driver with nvidia-smi first, then match CUDA version:

Reason 2: A Plugin Downgraded PyTorch

Some plugins' requirements replace your CUDA PyTorch with CPU-only. Fix: reinstall CUDA PyTorch, use --no-deps for risky plugins.

Reason 3: Wrong Batch File

Portable package: use run_nvidia_gpu.bat, not run_cpu.bat.

Reason 4: Driver Too Old

Update from nvidia.com.

Reason 5: RTX 50 Series Needs CUDA 12.8+

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130

See GPU Compatibility for the full matrix.

Verify

python -c "import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0))"

Should print True + your GPU name. Still False? Book a remote session.