Skip to content

[FEA] Reduce duplicate distance computations in Vamana GreedySearch build #2604

Description

@marcelo-cjl

Is your feature request related to a problem? Please describe.
The Vamana GPU build is dominated by GreedySearch on high-dimensional datasets. During each search, nodes reachable through more than one path are re-checked against the visited set and have their distances recomputed each time, which is redundant work that scales with graph degree.

Describe the solution you'd like
Reduce the per-search visited-set handling so each node's distance is computed at most once, lowering build latency for high-dimensional workloads without changing search results.

Describe alternatives you've considered
Keeping the current visited-list handling and accepting the redundant distance computations is the baseline; it is simple but leaves build time higher than necessary on high-dimensional data.

Additional context
Measured on NVIDIA L4 (graph_degree=64, visited_size=128): build time reduced 14-34% on sift/cohere/gist/openai with no change in recall@10.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions