Environment Variables

Various environment variables can be set to affect the behavior of the CUDA target.

NUMBA_CUDA_ARRAY_INTERFACE_SYNC

Whether to synchronize on streams provided by objects imported using the CUDA Array Interface. This defaults to 1. If set to 0, then no synchronization takes place, and the user of Numba CUDA MLIR (and other CUDA libraries) is responsible for ensuring correctness with respect to synchronization on streams.

NUMBA_CUDA_LOW_OCCUPANCY_WARNINGS

Enable warnings if the grid size is too small relative to the number of streaming multiprocessors (SM). This option is on by default (default value is 1).

The heuristic checked is whether gridsize < 2 * (number of SMs). NOTE: The absence of a warning does not imply a good gridsize relative to the number of SMs. Disabling this warning will reduce the number of CUDA API calls (during JIT compilation), as the heuristic needs to check the number of SMs available on the device in the current context.

NUMBA_CUDA_WARN_ON_IMPLICIT_COPY

Enable warnings if a kernel is launched with host memory which forces a copy to and from the device. This option is on by default (default value is 1).

NUMBA_CUDA_INCLUDE_PATH

The location of the CUDA include files. This is used when linking CUDA C++ sources to Python kernels, and needs to be correctly set for CUDA includes to be available to linked C/C++ sources. On Linux, it defaults to /usr/local/cuda/include. On Windows, the default is $env:CUDA_PATH\include.

NUMBA_CUDA_NVRTC_EXTRA_SEARCH_PATHS

A colon separated list of paths that NVRTC should search for when compiling external functions. These folders are searched after the system cudatoolkit search paths and Numba CUDA MLIR’s internal search paths.