Conversation
Pin cuda-pathfinder>=1.5 in both build-system.requires and project.dependencies. Made-with: Cursor
The previous 1.3.4a0 was a stale value that could confuse someone. Made-with: Cursor
Made-with: Cursor
Made-with: Cursor
…anning sys.path Made-with: Cursor
…17 builds" This reverts commit 3cb0313.
Use exclusion-style constraints in cuda_bindings, cuda_core, and cuda_pathfinder so CI avoids cuda-toolkit 12.9.2 and 13.0.3 while still allowing newer good patch releases. Made-with: Cursor
…ecs." This reverts commit 937ff7f.
|
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. |
3efe6e5 to
d6a5936
Compare
|
/ok to test |
|
|
This PR makes it so that we have to use to jump to our sources. One extra level, four characters more, I guess it's only a tiny extra inconvenience overall. But I'm wondering, could it also be: I guess not? — Or is there some place for the README, pyproject.toml etc. files with that layout? |
|
No. Different projects need to have their own src. I am supportive of this move, though my explicit complaint for the whole situation is: Python needs to do better in distinguishing local scripts/modules vs out-of-tree builds. Forcing everyone to move to |
Sounds good to me. Is there already a plan for moving cuda_core and cuda_pathfinder? (Doing all asap would seem best to me.) |
Yes, I agree. I think this is just an historical artifact that the language started as "run stuff out of the current directory" long before packages and virtual environments etc. were ever dreamed of. It still hasn't fully moved into the future.
I think we should do them all at the same time -- I just wanted to have a bot build a PoC first and prove it's possible. Mainly I wanted to confirm that none of the GHA would need to change, since that would indicate things our /users/ might have to change. There will be a little bit of fallout for anything that automates putting cython-gen/cybind code into this repo since the destination directory changes. |
This moves cuda_bindings to use a src layout, which is the recommended way to avoid a number of issues when the local checkout environment is getting inadvertently pulled in. While this may in fact be a bug in
pipthat it isn't sufficiently isolating the build environment, it's still good practice to usesrclayout, so we may want to consider it.Of course, a big change like this will have far-reaching merge implications and on other workflows etc., so I'm not suggesting we merge this as-is without some discussion.