Module index_builder 

GPU-accelerated HNSW index construction

This module implements GPU-based HNSW graph construction using CUDA kernels (feature-gated behind the cuda feature). When CUDA is not available, it falls back to an efficient CPU implementation.

Architecture

The GPU index builder works in phases:

1. Vector Upload: Transfer vectors to GPU memory in batches
2. Distance Matrix Computation: Compute all-pairs distances via CUDA kernels
3. Neighbor Selection: Apply heuristic neighbor selection on GPU
4. Graph Assembly: Assemble the HNSW graph structure from GPU results

Pure Rust Policy

All CUDA code is gated with #[cfg(feature = "cuda")] so the default build is 100% Pure Rust.

Structs

BatchSizeCalculator

Calculates optimal batch sizes for GPU index construction based on available GPU memory and vector dimensionality.

ComputedBatch

A batch for which GPU distance computation has been (simulated as) completed.

GpuBatchDistanceComputer

GPU-accelerated batch distance computation

GpuHnswIndexBuilder

GPU-accelerated HNSW index builder

GpuIndexBuildStats

Build statistics for the GPU index builder

GpuIndexBuilderConfig

Configuration for GPU-accelerated HNSW index building

GpuIndexOptimizer

Optimises GPU memory usage during index construction by computing ideal batch sizes and checking memory feasibility.

GpuMemoryBudget

GPU memory budget tracker for index construction.

HnswGraph

The assembled HNSW graph structure

HnswNode

A node in the HNSW graph

IncrementalGpuIndexBuilder

Incremental GPU index builder for streaming ingestion

IndexedBatch

A fully indexed batch: neighbor IDs have been selected and are ready to be merged into the final HNSW graph.

PipelinedIndexBuilder

Overlaps CPU preparation work with simulated GPU compute to build an index in a three-stage pipeline: prepare → compute → finalize.

PreparedBatch

A batch of vectors prepared (normalised / packed) on the CPU, ready to be dispatched to a GPU compute stage.

Enums

GpuDistanceMetric

Distance metric type for GPU kernels