#8279 CUDA GPU acceleration unavailable. Verify NVIDIA GPU & CUDA installation (Help -> CUDA Help).

CUDA acceleration is unavailable because the CUDA toolkit (12.8+), cuDNN 9.x runtime, or the NVIDIA driver on your system is missing, outdated, or not exposed to the program via PATH. The app can only enable CUDA when all three are present and visible.

1. Confirm your hardware is supported

• Only NVIDIA GPUs can use CUDA. AMD/Intel GPUs (integrated or discrete) cannot enable this option — for AMD you would need the Whisper.NET channel (source-deploy only).

2. Verify the installed CUDA version
Open CMD and run:

nvcc -V
nvidia-smi

• nvcc -V must report CUDA 12.8 or higher (12.8+, including 13.x).
• nvidia-smi shows your driver and the maximum CUDA version it supports. If the driver is old, update it from the NVIDIA website — an outdated driver alone will silently block CUDA.

3. Install / verify cuDNN 9.11+

• cuDNN must be 9.11 or newer.
• Add both the cuDNN bin and lib directories to the Windows Path environment variable. Missing bin on PATH is the most common cause of "CUDA unavailable" even when CUDA is installed.

4. Check environment variables

• The CUDA bin and lib folders (e.g. C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8\bin) must be in Path.
• After editing environment variables, fully close and restart the software (a restart of the machine is safer).

5. Additional notes for your setup

• Your install path D:/1. J.ALMANZA/win-PatchUpdate-v4.16 contains a space, a dot and mixed characters. Move the folder to a simple all-English path such as D:\pyVideoTrans — while this does not cause the CUDA error directly, non-trivial paths regularly cause secondary failures with bundled binaries.
• If you only applied the ~360 MB patch package on top of a very old version, re-download the full package (~2.7 GB) and overwrite, then delete every __pycache__ folder found inside _internal before retesting.
• If CUDA still cannot be enabled after all checks, run recognition with faster-whisper using int8 compute type on CPU, or switch to an online ASR channel.

Please refer to the documentation:

• Configuring CUDA and cuDNN on Windows
• faster-whisper (local built-in)
• FAQ