Vyas Ramasubramani

I probably won't have time to look into this further for another week or so, and I don't know this code all that well so it'll probably take me a...

Having skimmed the implementation here, I think that we need to take a step back here. What is the goal of the current design? In [`parallel_operation.hpp`](https://github.com/rapidsai/kvikio/blob/716a99f212ce2e0f9862bad1be81a12c56d03dc6/cpp/include/kvikio/parallel_operation.hpp) We are launching tasks...

OK, upon second read of the `parallel_operation.hpp` code I see what we're doing there. The future returned from `parallel_io` is just calling get on each contained future, but the actual...

I believe this was closed by #596

I'm fine doing this if we can get it working now. I forgot the plan from https://github.com/NVIDIA/cccl/issues/1939#issuecomment-2830993942 when I opened https://github.com/rapidsai/build-planning/issues/183, where I proposed waiting to do cccl until we...

OK makes sense. Now that we've started the CUDA 11 drops we'll need to make sure that we remember to debug this issue since we won't observe it in CI...

Is it possible/realistic to move all uses of nvtx out of rmm headers and into implementation files (as part of the move to rmm as a shared library)? Historically the...

Agreed, the templating makes this challenging. I don't know if it is feasible to completely avoid exporting some NVTX dependency to consumers, but I would like us to explore how...

I'd rather document this somewhere or have some snippets outside of the main library that we can look into this. It feels too specific to be something that actually lives...

Thanks for reporting this. From some quick digging, it looks like the issue here is that numpy's datetime object upcasts to an array and attempts to use the `__array_ufunc__` overload...