发布

  • Release: turbovec 0.5.1 (Python) + 0.4.1 (Rust crate)

    frostbyte_neo 发布于 2026-05-18 18:12:32 +00:00

    Performance release: block-level early exit for selective mask
    searches. Selective allowlists now short-circuit the SIMD kernel at
    32-vector block granularity (64-vector pair granularity on
    AVX-512BW). Closes #30.

    Measured speedup at 1% selectivity, 100K vectors, d=1536:

    • ARM (M3 Max): 0.98x -> 6.40x
    • x86 (Sapphire Rapids 8c): 0.98x -> 12.67x

    Speedup scales with index size — larger indices see larger
    speedups as the fixed per-query overhead (rotation, LUT build, top-k
    merge) amortizes.

    No change to existing public API. New Rust pub items for
    hybrid-retrieval telemetry: BLOCKS_SKIPPED_BY_MASK (atomic counter),
    blocks_skipped_by_mask(), reset_blocks_skipped_by_mask().

    No on-disk format change. Unmasked search latency unchanged.

    Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com

    下载附件