Single vector CAGRA index to not throw cudaErrorIllegalAddress (#666)

[chatman]

This is a fix for #666. It is useful for github.com/rapidsai/cuvs-lucene to avoid hacks that fallback to using Brute Force Index if dataset size is 1.

[copy-pr-bot]

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.

[achirkin]

Thanks, this is definitely something we should fix on the cuvs/C++ side. But also I think this deserves a deeper investigation: does this same error happen with just two vectors in the index? or three? Generally, how the graph is constructed for the number of vectors less than the graph degree?

[cjnolet]

Ishan, we should be using the tiered index in lucene so that we can avoid this very problem. Cagra graph is simply not valid for such small numbers of vectors. It’s meant to be a batched created index. It’s not the same we as HNSW. I’d prefer to focus the engineering team’s efforts on legitimate bugs and have cuVS-lucene use the proper index type.

…

On Thu, Aug 14, 2025 at 4:14 AM Artem M. Chirkin ***@***.***> wrote: *achirkin* left a comment (rapidsai/cuvs#1256) <#1256 (comment)> Thanks, this is definitely something we should fix on the cuvs/C++ side. But also I think this deserves a deeper investigation: does this same error happen with just two vectors in the index? or three? Generally, how the graph is constructed for the number of vectors less than the graph degree? — Reply to this email directly, view it on GitHub <#1256 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAJPKYEN6MQUQEKJ5BBP6WD3NRAPNAVCNFSM6AAAAACD3SCWOCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTCOBXGQYTIMBRG4> . You are receiving this because your review was requested.Message ID: ***@***.***>