Fix stream handle cast in Resources and MultiGpuResources by ViderStar · Pull Request #2946 · NVIDIA/cuvs

ViderStar · 2026-10-25T19:38:52Z

Resources(stream=...) and MultiGpuResources(stream=...) take the stream as an integer handle (the Resources docstring passes cupy.cuda.Stream().ptr). Both called

cuvsStreamSet(self.c_obj, <cudaStream_t>stream)

where stream is a Python object. Cython casts a Python object to a pointer type by taking the object's own address, so cuVS got the address of the Python int as the stream. The first call that used the stream then crashed. In the issue that call is res.sync().

The fix converts the value through uintptr_t first (<cudaStream_t><uintptr_t>stream) in both classes.

Tests: added python/cuvs/cuvs/tests/test_resources.py. It creates Resources from a CuPy stream handle, syncs it, and runs a small brute-force build and search on that handle.

How I tested (RTX 5070 Ti, CUDA 13 driver): I installed the cuvs-cu13 26.10 wheels, applied this change to their resources.pyx and mg_resources.pyx, and rebuilt the two modules with Cython against libcuvs_c. The pyx change is the same as on main. I could not build main's Python package against the 26.10 C library, because main uses newer C API functions.

Unmodified 26.10 modules: the issue's reproducer and the new test both segfault.
With the fix: the reproducer passes for both Resources and MultiGpuResources (single GPU), and test_resources.py, test_memory_tracking.py and test_brute_force.py all pass.

Resources(stream=...) takes the stream as an integer handle, for example cupy.cuda.Stream().ptr, but cast the Python object directly to cudaStream_t. Cython then passes the address of the Python int object instead of its value, so the handle got an invalid stream and sync() segfaulted. Convert through uintptr_t first. Signed-off-by: Artsem Lebiadzevich <artyomlebedevich@gmail.com>

copy-pr-bot · 2026-10-25T19:38:56Z

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.

ViderStar · 2026-10-25T19:39:02Z

Could a maintainer add the bug and non-breaking labels? I can't set labels from a fork.

coderabbitai · 2026-10-25T19:43:22Z

Important

Review skipped

Review was skipped as selected files did not have any reviewable changes.

Run configuration

Source mode: Auto
Repository: NVIDIA/cuvs
Branch: main
File: .coderabbit.yaml
Commit: 95601f8
Layers: Repository YAML
Review profile: CHILL
Plan: Enterprise
Run ID: bad83981-5a75-40a5-a21e-4a93c0ef892d

Commits

Reviewing files that changed from the base of the PR and between 04e1b6e and aa394dd.

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

No actionable comments were generated in the recent review.

Recent review info

Run configuration

Source mode: Auto
Repository: NVIDIA/cuvs
Branch: main
File: .coderabbit.yaml
Commit: 244f118
Layers: Repository YAML
Review profile: CHILL
Plan: Enterprise
Run ID: dcfc1602-89de-48dc-939e-a8891adf4c1a

Commits

Reviewing files that changed from the base of the PR and between 244f118 and d4f2dc4.

Files selected for processing: 3

python/cuvs/cuvs/common/mg_resources.pyx
python/cuvs/cuvs/common/resources.pyx
python/cuvs/cuvs/tests/test_resources.py

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 9 remain after this review.

📝 Summary

Summary by CodeRabbit

Bug Fixes
- Improved compatibility with externally provided GPU streams, including streams created by other Python GPU libraries. Index creation and search operations can now use these streams more reliably, while preserving the existing behavior when no stream is specified.
- Added coverage confirming that nearest-neighbor search returns the expected results when run using an external stream.

The Python Resources and MultiGpuResources constructors now convert truthy stream handles through uintptr_t before passing them to cuvsStreamSet. A test exercises Resources with a CuPy stream during brute-force index construction and search.

Changes: Python resource streams

Layer / File(s)	Summary
Stream conversion and external-stream test `python/cuvs/cuvs/common/mg_resources.pyx`, `python/cuvs/cuvs/common/resources.pyx`, `python/cuvs/cuvs/tests/test_resources.py`	Both constructors convert truthy stream handles through `uintptr_t` before casting them to `cudaStream_t`. The test synchronizes a `Resources` instance created with a CuPy stream, builds and searches a brute-force index, and checks the returned neighbors.

Priority: Normal

Estimated code review effort: 2 (Simple) | ~8 minutes

Change: Bug fix · Severity of issue fixed: Medium

Suggested labels: bug, non-breaking

Suggested reviewers: viclafargue, benfred, achirkin

Merge Risk: Minimal · previous review d4f2d

The reported stream-handle crash is addressed in both resource classes, with targeted external-stream coverage added.

Pre-merge checks: 5

Check name	Status	Explanation
Title check		The title names the main change: fixing the stream handle cast in the Resources and MultiGpuResources classes. The raw summary and objectives both describe this cast fix as the primary change, so the …
Description check		The description states the bug (the Python object address was passed instead of the integer handle value), the fix (converting through uintptr_t in both classes), the added test, and the testing perfo…
Linked Issues check		Issue `#1863` requires `Resources(stream=stream.ptr)` to synchronize without a segmentation fault. `resources.pyx` now converts the Python handle through `uintptr_t` before casting it to `cudaStream_t`.…
Out of Scope Changes check		The changes are limited to the related stream-handle conversion in `Resources` and `MultiGpuResources`, plus an automated regression test. The `MultiGpuResources` fix applies the same defect correctio…
Docstring Coverage		Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 1 functions across 1 files. (2 skipped: 2 …

Full details: Title check

Explanation

The title names the main change: fixing the stream handle cast in the Resources and MultiGpuResources classes. The raw summary and objectives both describe this cast fix as the primary change, so the title is specific and on-topic.

Full details: Description check

Explanation

The description states the bug (the Python object address was passed instead of the integer handle value), the fix (converting through uintptr_t in both classes), the added test, and the testing performed. It directly matches the changeset and linked issue #1863.

Full details: Linked Issues check

Explanation

Issue #1863 requires Resources(stream=stream.ptr) to synchronize without a segmentation fault. resources.pyx now converts the Python handle through uintptr_t before casting it to cudaStream_t. The new test_resources_with_external_stream creates a CuPy stream, constructs Resources with stream.ptr, calls res.sync(), and uses the resource for brute-force operations before synchronizing again.

Full details: Out of Scope Changes check

Explanation

The changes are limited to the related stream-handle conversion in Resources and MultiGpuResources, plus an automated regression test. The MultiGpuResources fix applies the same defect correction to the related class named in the PR intent. The test directly exercises the linked issue behavior and stream use.

Full details: Docstring Coverage

Explanation

Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 1 functions across 1 files. (2 skipped: 2 unsupported.)

Finishing Touches: 💡 1

🛠️ Fix failing CI checks: 💡

Commit to this branch
Create a new PR

🧪 Generate unit tests (beta)

Create a new PR

Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

…m-cast Signed-off-by: Artsem Lebiadzevich <artyomlebedevich@gmail.com>

ViderStar · 2026-10-28T17:32:34Z

@viclafargue @cjnolet gentle ping: could one of you review this when you have time? Passing a CUDA stream to the resources wrapper can segfault. I've merged the latest main. It also still needs the bug and non-breaking labels for the label checker.

viclafargue · 2026-10-28T17:37:39Z

/ok to test 04e1b6e

viclafargue

LGTM! Thanks for fixing the stream handle cast in both resource classes and adding a regression test.

viclafargue · 2026-10-28T17:45:41Z

/merge

aamijar · 2026-10-30T02:18:24Z

/ok to test aa394dd

aamijar

Nice! I remember opening that issue a while back

ViderStar requested a review from a team as a code owner October 25, 2026 19:38

github-project-automation Bot added this to Unstructured Data Processing Oct 25, 2026

Merge remote-tracking branch 'upstream/main' into fix-resources-strea…

04e1b6e

…m-cast Signed-off-by: Artsem Lebiadzevich <artyomlebedevich@gmail.com>

viclafargue self-requested a review October 28, 2026 17:36

viclafargue added bug Something isn't working non-breaking Introduces a non-breaking change labels Oct 28, 2026

viclafargue approved these changes Oct 28, 2026

View reviewed changes

github-project-automation Bot moved this to Approved in Unstructured Data Processing Oct 28, 2026

Merge branch 'main' into fix-resources-stream-cast

aa394dd

aamijar approved these changes Oct 30, 2026

View reviewed changes

rapids-bot Bot merged commit 25c22bd into NVIDIA:main Oct 30, 2026
71 checks passed

github-project-automation Bot moved this from Approved to Done in Unstructured Data Processing Oct 30, 2026

Navigation Menu

Repository navigation

Fix stream handle cast in Resources and MultiGpuResources - #2946

Fix stream handle cast in Resources and MultiGpuResources#2946
rapids-bot[bot] merged 3 commits into
NVIDIA:mainfrom
ViderStar:fix-resources-stream-cast

ViderStar commented Oct 25, 2026

Uh oh!

copy-pr-bot Bot commented Oct 25, 2026

Uh oh!

ViderStar commented Oct 25, 2026

Uh oh!

coderabbitai Bot commented Oct 25, 2026 •

edited

Loading

Review skipped

Summary by CodeRabbit

Uh oh!

ViderStar commented Oct 28, 2026

Uh oh!

viclafargue commented Oct 28, 2026

Uh oh!

viclafargue left a comment

Uh oh!

viclafargue commented Oct 28, 2026

Uh oh!

aamijar commented Oct 30, 2026

Uh oh!

aamijar left a comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

ViderStar commented Oct 25, 2026

Uh oh!

copy-pr-bot Bot commented Oct 25, 2026

Uh oh!

ViderStar commented Oct 25, 2026

Uh oh!

coderabbitai Bot commented Oct 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review skipped

Summary by CodeRabbit

Uh oh!

ViderStar commented Oct 28, 2026

Uh oh!

viclafargue commented Oct 28, 2026

Uh oh!

viclafargue left a comment

Choose a reason for hiding this comment

Uh oh!

viclafargue commented Oct 28, 2026

Uh oh!

aamijar commented Oct 30, 2026

Uh oh!

aamijar left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

coderabbitai Bot commented Oct 25, 2026 •

edited

Loading