Peter Andreas Entschev

Thanks for the ping @ian-r-rose . As you noted yourself, unfortunately CuPy Sparse is largely uncovered at the moment, I may be mistaken but I think it's usage is fairly...

> Apparently this was removed some time [ago last summer](https://github.com/rapidsai/rmm/issues/390) and was updated [in this PR](https://github.com/rapidsai/rmm/pull/626/) . @jrbourbeau please merge that PR when you feel it's ready and we'll fix...

I opened a PR with a fix in https://github.com/dask/distributed/pull/5158 .

> First, are you sure this is the right place to do this? When you say "this is the right place", are you referring to RMM itself, to the Cython...

> I mean the constructor of CudaStreamView. I suppose as long as there is still a way to explicitly request `cudaStreamLegacy` in this mode, it's OK. Sorry, this is still...

> In Numba we'd probably want to have an env var called `NUMBA_CUDA_PER_THREAD_DEFAULT_STREAM` because environment variables starting with `NUMBA_` are mapped into the `config` module such that the config variable...

> I mean that it already exists: pass `1` to the constructor of `CudaStreamView`. Although I don't like the magic numbers. I worked very hard in the C++ code to...

> I generally type env variables into scripts, not at the command line, so the length doesn't matter much. I do type variables, explicit is much better than implicit for...

Thanks @leofang , that's indeed very informative. I'm feeling less confused now, that was pretty much the only solution I could arrive at and your confirmation is quite helpful!

I changed `CudaStreamView` with the suggestions from https://github.com/rapidsai/rmm/pull/633#issuecomment-732323696, let me know if that's an acceptable solution.