Drop void return type when raising exceptions

[jakirkham]

Fixes #435

After poking at this a bit locally, it seems Cython's except * is picking up exceptions set through asyncio even when it shouldn't. Not sure if this behavior is due to Python or Cython. That said, we can avoid using except * here by using cdef without the void return type (this is the same as cdef object). Based on running things locally this seems to behave correctly for me. Though please check as well.

[jakirkham]

cc @quasiben @pentschev

[quasiben]

This still hangs for me with the dask-cuda merge benchmark. My Cython expertise is quite low but is there now conflict with argument and definitions in core_dep.pxd

ucx-py/ucp/_libs/core_dep.pxd

Line 136 in 9e3b3c0

ctypedef void (*ucp_send_callback_t)(void *request, ucs_status_t status) # noqa

?

[jakirkham]

Where are you running this? FWIW I ran this on dgx-15 without issues. Just want to exclude the possibility of hardware issues getting in the mix.

That's a good point. So I guess the exception will get ignored in those case anyways. Yeah it's hard to solve this correctly without more-or-less doing what Mads has already done. If we are not happy with this, maybe we can try to help get his work in? Sorry I don't have a better answer 😞

[jakirkham]

Should add when I tested this I was using Vibhu's MRE from PR ( #402 ). This seems potentially more complex than the merge benchmark. So I'm pretty sure I'm exercise all the same things, but please let me know if that's incorrect.

[pentschev]

Yeah, there's no error with the merge benchmark, Ben and I were testing both that and the merge benchmark. It's intriguing why it doesn't fail with that, in fact the tests for u1 and u16 fail, but u256 and larger pass, so it's probably due to the size of the transfer, but right now I have no clue where the issue may lie.

[madsbk]

This should be fixed in #418

[madsbk]

#439 replaces this PR