Repository navigation
v0.3.20
Improvements
Improve available-donor search with the built-in l2 and nan_euclidean metrics:
- Reuse buffers for distance corrections and shared-feature counts.
- Use floating-point matrix multiplication for overlap counting.
- Process search-preparation arrays in query groups and release temporary arrays before precise distance refinement.
- Add regression coverage for precision, chunk boundaries, cache reuse, and temporary-array lifetime.
Performance
A direct comparison of pre-release candidate b8e5a0f3 with PyPI 0.3.19 measured:
- 50–54% less first-transform time, excluding fit.
- 13.0–14.5% lower peak process RSS.
- Identical imputation outputs versus 0.3.19 in every measured case.
- All 108 benchmark workers passed their checks.
Measurements used 20,000 training rows, 300 queries, 20 features, one native thread, and nine workers per method and condition on an AMD EPYC 9V74.