cuda_core forward compatibility changes. #722
Conversation
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
|
/ok to test |
|
/ok to test |
This comment has been minimized.
This comment has been minimized.
|
/ok to test |
|
Auto-sync is disabled for ready for review pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
leofang
added
enhancement
| flags = 0 | ||
|
|
||
| status, _, graph, _, _ = handle_return(driver.cuStreamGetCaptureInfo(self._mnff.stream.handle)) | ||
| status, _, graph, *_, _ = handle_return(driver.cuStreamGetCaptureInfo(self._mnff.stream.handle)) |
There was a problem hiding this comment.
Q: If we're packing anyway, maybe we pack them together?
| status, _, graph, *_, _ = handle_return(driver.cuStreamGetCaptureInfo(self._mnff.stream.handle)) | |
| status, _, graph, *_ = handle_return(driver.cuStreamGetCaptureInfo(self._mnff.stream.handle)) |
|
I took a look at the graphs side and while I'm ok with it, it does feels over-engineered. Thinking out loud, what if we move these weird APIs to a compatibility file/layer? At least this way we're keeping these tricks contained in a single place. Oh and by tricks I mean logic like |
|
/ok to test |
I was very careful to make the unpacking consistent between all the uses, even where I could have taken a shortcut (
Yeah, it's a bit over the top, but there are only two, and the circumstances are special. |
|
Thanks, Ralf! |
|
3 participants
Description
Forward compatibility changes for
cuda_core.This helps us staying organized.