Graham Markall

> @pentschev I'm a bit confused as to how you got that error. I just saw https://github.com/rapidsai/dask-cuda/pull/1201, and now I think maybe the error came when the CUDA toolkit wasn't...

Perhaps you have `cuda-python` 12 installed, which is looking for NVRTC from CUDA 12. I think you could try either upgrading your CUDA toolkit to 12 or downgrading `cuda-python` to...

BTW if you don't want to rely on cuda-python, you could apply https://github.com/numba/numba/pull/9086 locally, which should make it into 0.58.

> After further exploration, we have found all the files above can be gotten with just 2 dependencies on CUDA 12: > > * `cuda-nvcc-impl` > > * `cuda-nvrtc` >...

Coming back to look at this, I realise I don't recall what things like "NB_NPY20" mean - can we add a key that explains what each item is, for the...

This is probably worth re-visiting following the 0.59 release (which is a priority and consuming all available time right now).

This may be a good issue for someone who's already managed to implement a basic addition to the CUDA target - labelling accordingly.

A little further investigation reveals that `numba_attempt_nocopy_reshape` is implemented in `_helperlib.c` - this could do with being compiled to PTX into a "CUDA helperlib" and linked in as required. This...

Starter hack: `_cuda_helperlib.cu`, taken from `_numba/_helperlib.c`: ```cuda #define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION #include /* * Handle reshaping of zero-sized array. * See numba_attempt_nocopy_reshape() below. */ extern "C" __device__ int nocopy_empty_reshape(npy_intp nd, const...

xref #2463 which should be a use case enabled by implementing this.