You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The NVIDIA CUDA Python bindings are required to link C/C++ code with Numba kernels, because it compiles them with NVRTC. This means that full float16 support also needs those bindings. If we supported NVRTC with the ctypes bindings, then float16 support would always be available.
Previously this was not implemented because it might have been tricky to locate NVRTC, as it wasn't present in cudatoolkit conda packages. However, from CUDA 12 onwards, the cudatoolkit (or cuda-toolkit?) packages on conda-forge will have it available, so locating it should be less of an issue.
The text was updated successfully, but these errors were encountered:
The NVIDIA CUDA Python bindings are required to link C/C++ code with Numba kernels, because it compiles them with NVRTC. This means that full
float16
support also needs those bindings. If we supported NVRTC with the ctypes bindings, thenfloat16
support would always be available.Previously this was not implemented because it might have been tricky to locate NVRTC, as it wasn't present in
cudatoolkit
conda packages. However, from CUDA 12 onwards, thecudatoolkit
(orcuda-toolkit
?) packages on conda-forge will have it available, so locating it should be less of an issue.The text was updated successfully, but these errors were encountered: