Updating pylibraft pairwise_distance to cuvs#8847
Conversation
…ards compatible with pylibraft when appropriate)
Co-authored-by: Bradley Dice <[email protected]>
|
I assume these two files will need to change too, in order to test this in the CI: |
|
Thanks @cjnolet, LGTM! Could you bump the base Docker image to the one with cuVS (I guess |
|
/test rapids |
|
/test rapids |
1 similar comment
|
/test rapids |
|
/test rapids |
|
/test rapids |
|
/test rapids |
|
@kmaehashi I'm confused about an error I'm getting in CI: It started happening right after updating the link to the docker image (which uses 12.5 for all cuda 12 versions), but I'm not sure if it's related or if just a coincidence. |
|
/test rapids |
|
/test rapids |
|
/test rapids |
|
/test rapids |
|
/test rapids |
|
Okay, I've got CI running successfully and all tests are passing. Last piece here is waiting unti this cuVS PR is merged and the nightly is deployed: NVIDIA/cuvs#570. Then we can merge this PR. |
|
/test rapids |
|
/test rapids |
|
@kmaehashi this is ready to merge whenever you are ready. Thanks so much! |
|
@kmaehashi Shall we merge this PR ? |
|
/merge |
|
@kmaehashi just a friendly ping. I don't think I have the power to merge this PR. Any chance we could get it merged? |
|
Sorry for the delay, thanks again! |
|
@kmaehashi Failed to backport automatically. |
|
I'll take care of backport to v13 |
Updating pylibraft `pairwise_distance` to cuvs
This PR updates cupy so that it can make use of the cuvs library for pairwise distance, which the vector search APIs from pylibraft have recently moved. This PR maintains the backwards compatibility with pylibraft while will exposing double precision float support.
Closes NVIDIA/cuvs#542