GPU error fixes

The errors people hit when PyTorch or TensorFlow meets a GPU, each answered from this site's data: what every build was compiled for, what every driver supports. Not generic advice.

no kernel image is available for execution on the device

PyTorch installed, but the build has no code for your GPU. Most common on older NVIDIA cards (Pascal, Volta) with the default install.

Also covers: CUDA capability sm_XX is not compatible with the current PyTorch installation

CUDA driver version is insufficient for CUDA runtime version

Your driver is older than the CUDA build needs. Includes how to read the "found version" number.

Also covers: The NVIDIA driver on your system is too old (found version …)

torch.cuda.is_available() returns False

Usually a CPU-only build (the Windows default), otherwise the driver, a container, WSL2 or AMD permissions.

Also covers: Torch not compiled with CUDA enabled · Found no NVIDIA driver on your system

hipErrorNoBinaryForGpu: Unable to find code object for all current devices!

AMD: the build has no code for your card's gfx target. Which PyTorch ROCm build covers each Radeon and Instinct card.

Also covers: HIP error: invalid device function

Could not load dynamic library 'libcudart.so…'

TensorFlow can’t find CUDA and falls back to the CPU. A one-line fix from TensorFlow 2.14 on.

Also covers: Cannot dlopen some GPU libraries

For a specific GPU, driver and Python version, the install checker gives the exact install command.