Skip to content

Releases: M4iKZ/Vector-Arena

Vector-Arena RC1: Optimized Custom Engines & High-Performance Benchmark Suite

06 Apr 06:08
@M4iKZ M4iKZ
RC1
b7c0b7c

Choose a tag to compare

View all tags
Vector-Arena RC1: Optimized Custom Engines & High-Performance Benchmark Suite Pre-release
Pre-release

Overview

Vector-Arena RC release 1 officially introduces the integration of two high-performance custom engines: mSEARCH and MeMo. This release includes precompiled binaries (.pyd) specifically built for Python 3.13, allowing immediate benchmarking of these engines.

🛠 Custom Engine Features (Included Binaries)

  • MeMo (pymemo_embedded.cp313-win_amd64.pyd): High-performance embedded database with zero-copy document ingestion and a fluent query API.
  • mSEARCH (msearch.cp313-win_amd64.pyd): Modular HNSW vector search implementation.
    • SIMD Optimized: Native AVX2/FMA/NEON acceleration for Windows.
    • Native FP16: Optimized half-precision support for reduced memory usage.

📥 Using the Windows Binaries (Requires Python 3.13)

To enable the custom engines, download the attached .pyd files and place them into the numerable folder located at the root of the benchmark directory.

Important

These binaries are compiled specifically for Python 3.13. They will not work if you are using an older Python version (e.g., 3.12 or 3.11).

📊 Benchmark Suite Refactor

  • Multi-Dataset Support: Standardized benchmarking for sift, glove-25, glove-100, deep1b, nytimes, and fashion-mnist.
  • Dynamic Recall Validation: Accurately calculates recall for subsets (1k, 10k, 100k), ensuring valid precision metrics at any scale.
  • Subprocess Isolation: Each engine runs in a separate process for clean resource tracking (RSS/Memory) and crash isolation.

📈 Reporting & Visualization

  • Automated Reports: Generates detailed GitHub-flavored Markdown summaries.

Getting Started

  1. Ensure you are running Python 3.13.
  2. Install requirements: pip install -r vector_bench_requirements.txt
  3. Download the precompiled .pyd binaries from this release and move them into the ./numerable/ folder.
  4. Run the benchmark: python vector_bench.py --docs 10000 --datasets sift,glove-25
Loading