Graham Markall

In the long term the changes will be incorporated into Numba's CUDA target, so that pynvjitlink is not needed for patching the Numba linker.

@ed-o-saurus Sorry I forgot to follow up earlier - the functionality is available in numba-cuda, but changes will not be backported to the built-in target - there is some explanation...

I'd been meaning to look into Enzyme, as it also seems to support CUDA-specific techniques. @adam-urbanczyk have you any thoughts or experiments in the direction of AD support for Numba...

I think this is an area none of the Numba maintainers have any significant amount of experience with, so it'll be difficult for us to push forward on this without...

> @gmarkall do I understand it correctly that Numba does not use CUDA-Streams, or CUDA-Graphs for its CUDA support? Apologies for the delay in my reply here. It doesn't use...

> @gmarkall coming back to the original question, is the description good enough? Do you need more details? I guess so - as my experience in this area is extremely...

> Hi :) Giving an updated talk next week at EuroAD, will gladly post the slides after the talk Slides appear to be here: https://www.autodiff.org/Docs/euroad/25th%20EuroAd%20Workshop%20-%20Ludger%20Paehler%20-%20Numba-Enzyme:%20Differentiable%20JIT%208217%20d%20Python.pdf

I'm afraid there is no update at the moment.

A belated thanks for this PR - rather than attempting to go our own way in Numba CUDA (to avoid duplication of effort / minimise the amount of additional bespoke...

After merging master into this branch there is a small conflict which is trivial to resolve, and the added tests of `np.tile` pass - @AndrewEckart would you like to merge...