Ensure allocation stream is used for buffer deallocation if no explicit stream is provided by leofang · Pull Request #1032 · NVIDIA/cuda-python

leofang · 2025-09-26T05:07:51Z

Description

closes#779.

Checklist

New or existing tests cover these changes.
The documentation is up to date with these changes.

copy-pr-bot · 2025-09-26T05:07:55Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

leofang · 2025-09-26T14:14:08Z

/ok to test 738eed4

…ng warning

leofang · 2025-10-05T21:22:33Z

(This PR is now based entirely on #1070 except for the last commit, so that PR needs to be merged first.)

leofang · 2025-10-05T21:22:50Z

/ok to test 605241e

kkraus14

Is there any way we could add some testing for stream ordered allocations / deallocations?

kkraus14 · 2025-10-07T02:32:33Z

Changes LGTM on top of #1070

leofang · 2025-10-07T19:28:53Z

Merged with main!

leofang · 2025-10-07T19:29:04Z

/ok to test 33345d6

rparolin · 2025-10-07T21:51:16Z

+ If the buffer is deallocated without an explicit stream, the allocation stream
+ is used.
 """
- if stream is None:


Did we intend to remove this? Based on the deallocate API signature a user would deem it valid to pass None as a stream argument. That None is not checked in the _deallocate function above when the stream object is called with it.

This is a very interesting question that's worth expanding.. From my perspective mr.deallocate() is not meant to be called by end users because
it was designed to be called by Buffer.close() or its destructor when out of scope
it has such an ugly, not pythonic signature (passing pointer and size)
I don't think currently we have any test that actually exercise calling mr.deallocate() explicitly.
If this can be called by the end users, the contract that we deallocate using the allocation stream would not be upheld, because this API does not take any Buffer argument and so the stream is not available to us.

leofang · 2025-10-07T22:34:00Z

Thanks, all! I noticed I forgot to prepare a release note for this PR. Working on it now...

leofang · 2025-10-07T22:39:16Z

Working on it now...

#1097

github-actions · 2025-10-07T22:46:44Z

Doc Preview CI
Preview removed because the pull request was closed or merged.

leofang self-assigned this Sep 26, 2025

leofang added P0 High priority - Must do! cuda.core Everything related to the cuda.core module breaking Breaking changes are introduced labels Sep 26, 2025

leofang added this to the cuda.core beta 7 milestone Sep 26, 2025

This comment has been minimized.
Sign in to view

cpcloud reviewed Sep 30, 2025
View reviewed changes

Comment threadcuda_core/cuda/core/experimental/_memory.pyx Outdated

cpcloud reviewed Sep 30, 2025
View reviewed changes

Comment threadcuda_core/cuda/core/experimental/_memory.pyx Outdated

leofang added 17 commits October 3, 2025 03:34

release gil
1f84a5a

release gil for record
2c86825

use __dealloc__ in event/stream
efe7fe0

nit: remove print
fa88b28

fix linter error
1083d90

reduce further the number of Python objects held by Stream
c06dc9d

replace a few more __del__ by __dealloc__
b713c5c

improve __cuda_stream__ performance
d764f5b

cythonize Buffer & DMR (WIP - not working!)
3684644

minor fixes - still failing
2c9d5ec

fix casting and avoid repetive assignment
638bf59

ensure we have C access for DeviceMemoryResource
0c64e8e

restore the contracts for now (the deallocation stream should be fixed)
4c4bb17

fully cythonize Stream
c593ac9

make linter happy
fae5e7f

Merge branch 'main' into cythonize_more
bf712bf

ensure allocation stream can be used for deallocation; suppress casti…
605241e
…ng warning

leofang force-pushed the dealloc_stream branch from 738eed4 to 605241eCompare October 5, 2025 21:20

leofang added the blocked This task is currently blocked by other tasks label Oct 5, 2025

leofang requested a review from shwina October 6, 2025 13:13

kkraus14 reviewed Oct 6, 2025
View reviewed changes

Comment threadcuda_core/cuda/core/experimental/_memory.pyx Outdated

leofang mentioned this pull request Oct 6, 2025
Cythonize cuda.core more #1070
Merged
2 tasks

rwgk mentioned this pull request Oct 6, 2025
supported_nvidia_libs.py updates: add nvidia-cublasmp-cu12, nvidia-cublasmp-cu13, nvidia-cudss-cu13 #1089
Merged

int -> intptr_t -> void* is safer
43345d6

leofang removed the blocked This task is currently blocked by other tasks label Oct 7, 2025

Merge branch 'main' into dealloc_stream2
33345d6

leofang marked this pull request as ready for review October 7, 2025 19:28

leofang requested review from kkraus14 and rparolin October 7, 2025 20:01

rparolin reviewed Oct 7, 2025
View reviewed changes

Comment threadcuda_core/cuda/core/experimental/_memory.pyx

rparolin reviewed Oct 7, 2025
View reviewed changes

Comment threadcuda_core/cuda/core/experimental/_memory.pyx

rparolin approved these changes Oct 7, 2025
View reviewed changes

leofang merged commit 4cbb627 into NVIDIA:mainOct 7, 2025
71 checks passed

leofang deleted the dealloc_stream branch October 7, 2025 22:33

leofang mentioned this pull request Nov 7, 2025
Add a "known issues & limitations" page #1223
Open

Uh oh!

Conversation

leofang commented Sep 26, 2025

Description

Checklist

Uh oh!

copy-pr-botBot commented Sep 26, 2025

Uh oh!

leofang commented Sep 26, 2025

Uh oh!

This comment has been minimized.

Uh oh!

Uh oh!

leofang commented Oct 5, 2025

Uh oh!

leofang commented Oct 5, 2025

Uh oh!

kkraus14 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

kkraus14 commented Oct 7, 2025

Uh oh!

leofang commented Oct 7, 2025

Uh oh!

leofang commented Oct 7, 2025

Uh oh!

Uh oh!

rparolinOct 7, 2025

Choose a reason for hiding this comment

Uh oh!

leofangOct 7, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

leofang commented Oct 7, 2025

Uh oh!

leofang commented Oct 7, 2025

Uh oh!

github-actionsBot commented Oct 7, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants