User request (to CuPy) for setting cudaFuncAttributeNonPortableClusterSizeAllowed: cupy/cupy#8851

I think we should provide both avenues for specifying clusters or other configs which can be done via the kernel attributes of launch configs. There are 3 reasons for this:

1. Runtime vs compile time means they're not exactly the same

2. A user coming from c++ will expect both options, and a new user doesn't need to understand both options at first in order to get started. The possibility of being confused by two options is outweighed by the potential confusion of learning about both options via existing cuda c++ documentation or looking at the source code, but not being able to use them.

3. For us to provide both requires almost no effort, as we are leveraging the existing backend. Isolating things like clusters to one of either lunch configs, or kernel attributes, would require us to hack away a subset of perfectly valid attributes/configs on either launches or kernels.

I think both are run-time configs and if a kernel has any attribute that can be configured at (JIT-)compile time, IIRC there would not be any setter (I could be wrong). Another note is the driver team seems to have a preference over one of them. We should check with them internally for guidance.

I will get in touch, is there anyone in particular that comes to mind for asking about driver team preference?

I've got a review in flight. I never found (I didn't look very hard) documentation for which attributes were read vs write, but it was easy to deduce via testing. #360

is there anyone in particular that comes to mind for asking about driver team preference?

@vzhurba01 might know who on the driver team to discuss about this. Let's follow up offline.