Leo Fang

Thanks, @njzjz, so what I asked Cliff above would not work...

This needs: 1. a short example added to Tutorial 2. docstrings added to the relevant classes (ex: `Request`)

@dalcinl If IIUC non-default instances of the following 9 (*) classes need an explicit destructor call: - Datatype - Win - Request - Op - Info - Group - Comm...

My old PR that @dalcinl referred to is here: https://bitbucket.org/mpi4py/mpi4py/pull-requests/18. Feel free to work based on it and send a PR (to GitHub; we have abandoned Bitbucket) as I am...

By the way, forgot to reply: > Maybe something like a `gpu_support` object which has properties if a given API is supported and then that is set to either true...

My (conflicting) thoughts: - Maybe we could store an `is_cpu_accessible` flag in the [`internal`](https://docs.python.org/3/c-api/buffer.html#c.Py_buffer.internal) field of the buffer to check if we should disable accessing GPUs? - While mpi4py does...

I have been thinking about this but I'm still unsure if there's a robust way for handling this. Specifically, how do I reliably tell if MS-MPI is from Conda or...

Yes, this is exactly what I was referring to. Don't we *always* look up from the Windows registry (by calling `msmpi_ver()`)? https://github.com/mpi4py/mpi4py/blob/18f16da45eaaacfba4cc0293964e1a6548d4c6a6/conf/mpiconfig.py#L193-L198 It shouldn't be necessary to either look up...

> Do you agree that defining `MSMPI_VER` is ultimately harmless if building against the conda-forge package? Well, apparently it's harmless only because we forgot to set `-Werror` in mpi4py-feedstock. See,...

> Yes, of course. But I would not say we "forgot" to add it. IMHO, `-Werror` is not something you would add in production. This I totally I agree. Though...