[WIP] IP Vector Normalization to avoid all vectors clumped into single cluster in IVF-PQ by HowardHuang1 · Pull Request #1892 · rapidsai/cuvs

Commits on Mar 16, 2026

Commits on Mar 18, 2026

Commits on Mar 19, 2026

Commits on Mar 20, 2026

Commits on Mar 23, 2026

fix cuvs-bench generate groundtruth which was sorted incorrectly for inner product
HowardHuang1
committed

Commits on Mar 30, 2026

revert normalization outside kernel that requires copy of dataset resulting in 2x memory spike
HowardHuang1
committed

Commits on Mar 31, 2026

didn't modify kernels themselves but routed IP predict through existing fused cosine implementation
HowardHuang1
committed

Commits on Apr 2, 2026

add inner_product_cosine_assignment flag which is set when ivf_pq_build is called because only ivf_pq should have normalization for IP, other index types should not --> this version fixes low recal…
HowardHuang1
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[WIP] IP Vector Normalization to avoid all vectors clumped into single cluster in IVF-PQ #1892

[WIP] IP Vector Normalization to avoid all vectors clumped into single cluster in IVF-PQ #1892
HowardHuang1 wants to merge 21 commits into
rapidsai:mainfrom
HowardHuang1:HH-Vector-Normalization

Commits on Mar 7, 2026

Fixed vector normalization for Inner Product -- still failing 15 test cases due to reduced precision 8-bit values stored in LUT

Commits on Mar 9, 2026

Merge branch 'main' into HH-Vector-Normalization

Commits on Mar 10, 2026

add logging to compare recall and search speed for raw v.s. normalized IVF-PQ IP

Commits on Mar 11, 2026

revert change comparing normalize and raw in single run -- that was previously for debug only

Commits on Mar 16, 2026

previously normalization was applied to entire IVF-PQ pipeline --> change to only apply normalization to kmeans step

revert to raw vectors. No need to normalize here because normalized v.s. raw doesn't affect ordering of clusters for nearest cluster assignment

clean up code

clean up code

Commits on Mar 18, 2026

upload code that resolves linker issue + live csv updates

remove live_csv

Merge branch 'main' into HH-Vector-Normalization

Commits on Mar 19, 2026

hardcode file path instead of searching multiple directories + fix input command parsing

clean up unnecessary checks in data_export.py

Commits on Mar 20, 2026

bring back comma parsing instead of underscore parsing

bring back parts of plot/main.py for clarity

get rid of incremental JSON->CSV write for clarity

bring back original plot/main.py for clarity

Commits on Mar 23, 2026

fix cuvs-bench generate groundtruth which was sorted incorrectly for inner product

Commits on Mar 30, 2026

revert normalization outside kernel that requires copy of dataset resulting in 2x memory spike

Commits on Mar 31, 2026

didn't modify kernels themselves but routed IP predict through existing fused cosine implementation

Commits on Apr 2, 2026

add inner_product_cosine_assignment flag which is set when ivf_pq_build is called because only ivf_pq should have normalization for IP, other index types should not --> this version fixes low recal…