Skip to content

Faulty comms in Dask kNN #7573

@viclafargue

Description

@viclafargue

Since ucx-py was removed from the environment, Dask NearestNeighbors, KNeighborsClassifier and KNeighborsRegressor started to produce gibberish outputs resulting in low accuracy in tests. This was caused by issues in the RAFT communicator and is probably related to the removal of the UCX-py backend in favor of UCXX. Even though the UCX build in the environment is CUDA-aware, only the communications with pointers on host memory were transmitted successfully.

Since these estimators only exchange device memory, bypassing UCX and using the NCCL library directly (still through the RAFT communicator though) would solve the issue at least temporarily.

Metadata

Metadata

Assignees

Labels

bugSomething isn't working

Type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions