Skip to content

Conversation

@aamijar
Copy link
Member

@aamijar aamijar commented Sep 16, 2025

Resolves #7225; Depends on rapidsai/raft#2793 and rapidsai/cuvs#1302

For larger datasets UMAP dispatches a COO matrix with uint64_t nnz template types. This PR add support to correctly forward this template type to spectral initialization.

@copy-pr-bot
Copy link

copy-pr-bot bot commented Sep 16, 2025

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.

@aamijar aamijar added non-breaking Non-breaking change improvement Improvement / enhancement to an existing function algo: umap labels Sep 16, 2025
aamijar and others added 3 commits September 17, 2025 23:41
This PR changes two things:
1. Removes dependency on `libcuvs` for wheel builds, as we clone and build from source to facilitate static link
2. Removes dependency on `cuvs` for all builds

Authors:
  - Divye Gala (https://github.com/divyegala)
  - Kyle Edwards (https://github.com/KyleFromNVIDIA)

Approvers:
  - Kyle Edwards (https://github.com/KyleFromNVIDIA)
  - Simon Adorf (https://github.com/csadorf)

URL: rapidsai#7261
@csadorf
Copy link
Contributor

csadorf commented Sep 25, 2025

@aamijar Should this be moved to 25.12?

@aamijar
Copy link
Member Author

aamijar commented Sep 25, 2025

Yes, I guess the github issue was in 25.12, but the PR still says 25.10. Changed it now

aamijar and others added 4 commits September 25, 2025 11:29
Closes rapidsai#6539
~Requires rapidsai/raft#2739

This PR:
- Trims the graph before embedding initialization
- Stores the graph on host when using the UMAP estimator

Authors:
  - Victor Lafargue (https://github.com/viclafargue)

Approvers:
  - Jinsol Park (https://github.com/jinsolp)
  - Divye Gala (https://github.com/divyegala)

URL: rapidsai#6595
This makes use of the `CUML_ACCEL_ENABLED` environment variable to ensure if cuml.accel is enabled in a parent process then it's enabled (by default) in a child process. This makes `cuml.accel` usage more transparent with common python multiprocessing solutions (`multiprocessing`/`joblib`/`dask`). Things _just work_.

Currently on top of rapidsai#7046. Note that this use case was the primary motivation for rapidsai#7046, though I can see other reasons why an environment variable might be nice too.

Authors:
  - Jim Crist-Harif (https://github.com/jcrist)
  - Simon Adorf (https://github.com/csadorf)

Approvers:
  - Simon Adorf (https://github.com/csadorf)

URL: rapidsai#7047
@csadorf
Copy link
Contributor

csadorf commented Sep 27, 2025

@aamijar This is still targeting the 25.10 branch.

@aamijar aamijar changed the base branch from branch-25.10 to branch-25.12 September 27, 2025 01:28
@aamijar
Copy link
Member Author

aamijar commented Sep 27, 2025

Changed target branch now

@csadorf csadorf linked an issue Nov 17, 2025 that may be closed by this pull request
@aamijar
Copy link
Member Author

aamijar commented Nov 19, 2025

Approach has been updated, new PR will be created for this one.

@aamijar aamijar closed this Nov 19, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

algo: spectral-embedding algo: umap CUDA/C++ improvement Improvement / enhancement to an existing function non-breaking Non-breaking change

Projects

None yet

Development

Successfully merging this pull request may close these issues.

UMAP spectral initialization with int64_t nnz types. [BUG] CUDA_ERROR_ILLEGAL_ADDRESS issue with cuML UMAP

5 participants