Skip to content

Backend

This section documents backend-related utilities exposed by KeyDNN’s public API.

Backend APIs allow users to query system capabilities without directly depending on internal infrastructure or native bindings.

cuda_available

Return whether KeyDNN's CUDA native backend appears to be available.

This function attempts to load the KeyDNN CUDA native library via load_keydnn_cuda_native(). If the library loads successfully, CUDA is considered available; otherwise, the function returns False.

The check is best used as a fast feature gate for optional CUDA execution.

Returns:

Type Description
bool

True if the native CUDA backend can be loaded in the current process; False otherwise.
Notes
  • All exceptions raised by the underlying loader are swallowed and mapped to False. If you need diagnostics, call load_keydnn_cuda_native() directly to surface the underlying error.

Notes

  • cuda_available() can be used to conditionally enable CUDA-specific logic.
  • This function reflects whether CUDA support is available at runtime, not whether a particular tensor or model is currently on a CUDA device.
  • Backend initialization is handled internally; users should not manually load or manage native libraries.