Fix persistent CAGRA regressions#1800
Merged
Merged
Conversation
achirkin
added
bug
cjnolet
approved these changes
Feb 13, 2026
Contributor
|
/ok to test 1e45d9f |
Contributor
Author
|
/merge |
github-project-automation
Bot
moved this from In Progress
to Done
in Unstructured Data Processing
rapids-bot Bot
pushed a commit
that referenced
this pull request
May 9, 2026
CUDA 13 binary size reduction from 282 MB to 257 MB (-8.86%). Benchmark: <img width="1137" height="776" alt="image" src="https://github.com/user-attachments/assets/81b4f25a-c999-423d-a2c9-3e45566904bf" /> Apply updates from CAGRA related PRs: - [x] #1800 - [ ] #1781 - [x] #1780 - [x] #1771 - [x] #1834 - [x] #1851 - [x] #1831 JIT related PRs: - [x] #1812 Authors: - Divye Gala (https://github.com/divyegala) - Kyle Edwards (https://github.com/KyleFromNVIDIA) - Dante Gama Dessavre (https://github.com/dantegd) - Bradley Dice (https://github.com/bdice) Approvers: - Kyle Edwards (https://github.com/KyleFromNVIDIA) - Dante Gama Dessavre (https://github.com/dantegd) URL: #1807
landrumb
pushed a commit
to landrumb/cuvs
that referenced
this pull request
Jul 8, 2026
Fix peristent cagra kernel bugs: 1. Wrong arguments passed to the persistent kernel NVIDIA#1498 . Didn't catch this at compile time, because all affected parameters are of the same type. Never crashed in tests because of a lucky parameter combinations. But did crash in the example code. 2. Some changes in RMM over multiple versions made the pool allocator a bad choice for high-concurrency code, leading to the non-persistent async kernel call loop swap the CPU and hang forever. 3. Improved the speed of both implementations in the async loop by reducing the thread-launch overhead. Before: exactly `work_size` were spawned within a loop, competely dominating the runtime. After: `n_threads` are spawned and run an inner work loop over their shares of the input data. `n_threads` here is optimal for each implementation (persistent kernel requiring many more threads). The launch overheads are still significant, but do not overwhelm the useful work times. Authors: - Artem M. Chirkin (https://github.com/achirkin) Approvers: - Corey J. Nolet (https://github.com/cjnolet) URL: NVIDIA#1800
landrumb
pushed a commit
to landrumb/cuvs
that referenced
this pull request
Jul 8, 2026
CUDA 13 binary size reduction from 282 MB to 257 MB (-8.86%). Benchmark: <img width="1137" height="776" alt="image" src="https://github.com/user-attachments/assets/81b4f25a-c999-423d-a2c9-3e45566904bf" /> Apply updates from CAGRA related PRs: - [x] NVIDIA#1800 - [ ] NVIDIA#1781 - [x] NVIDIA#1780 - [x] NVIDIA#1771 - [x] NVIDIA#1834 - [x] NVIDIA#1851 - [x] NVIDIA#1831 JIT related PRs: - [x] NVIDIA#1812 Authors: - Divye Gala (https://github.com/divyegala) - Kyle Edwards (https://github.com/KyleFromNVIDIA) - Dante Gama Dessavre (https://github.com/dantegd) - Bradley Dice (https://github.com/bdice) Approvers: - Kyle Edwards (https://github.com/KyleFromNVIDIA) - Dante Gama Dessavre (https://github.com/dantegd) URL: NVIDIA#1807
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fix peristent cagra kernel bugs:
work_sizewere spawned within a loop, competely dominating the runtime. After:n_threadsare spawned and run an inner work loop over their shares of the input data.n_threadshere is optimal for each implementation (persistent kernel requiring many more threads). The launch overheads are still significant, but do not overwhelm the useful work times.