Module gpu 

GPU acceleration for vector operations using CUDA

This module provides GPU acceleration for:

• Distance calculations (cosine, euclidean, etc.)
• Batch vector operations
• Parallel search algorithms
• Matrix operations for embeddings

CUDA Feature Gating (Pure Rust Policy)

GPU acceleration is optional and properly feature-gated:

• Default build: 100% Pure Rust, no CUDA required, CPU implementations only
• With cuda feature: GPU acceleration when CUDA toolkit is installed
• With cuda feature but no toolkit: Graceful fallback to CPU implementations

All CUDA-dependent code is gated with #[cfg(all(feature = "cuda", cuda_runtime_available))] to ensure the crate builds successfully regardless of CUDA availability.

Re-exports

pub use accelerator::create_default_accelerator;

pub use accelerator::create_memory_optimized_accelerator;

pub use accelerator::create_performance_accelerator;

pub use accelerator::is_gpu_available;

pub use accelerator::GpuAccelerator;

pub use buffer::GpuBuffer;

pub use config::GpuConfig;

pub use config::OptimizationLevel;

pub use config::PrecisionMode;

pub use device::GpuDevice;

pub use index::AdvancedGpuVectorIndex;

pub use index::BatchVectorProcessor;

pub use index::GpuVectorIndex;

pub use index_builder::BatchSizeCalculator;

pub use index_builder::ComputedBatch;

pub use index_builder::GpuBatchDistanceComputer;

pub use index_builder::GpuDistanceMetric;

pub use index_builder::GpuHnswIndexBuilder;

pub use index_builder::GpuIndexBuildStats;

pub use index_builder::GpuIndexBuilderConfig;

pub use index_builder::GpuIndexOptimizer;

pub use index_builder::GpuMemoryBudget;

pub use index_builder::HnswGraph;

pub use index_builder::HnswNode;

pub use index_builder::IncrementalGpuIndexBuilder;

pub use index_builder::IndexedBatch;

pub use index_builder::PipelinedIndexBuilder;

pub use index_builder::PreparedBatch;

pub use load_balancer::GpuLoadBalancer;

pub use load_balancer::SimpleGpuDevice;

pub use load_balancer::WorkloadChunk;

pub use load_balancer::WorkloadDistributor;

pub use memory_pool::GpuMemoryPool;

pub use multi_gpu::GpuDeviceMetrics;

pub use multi_gpu::GpuTaskOutput;

pub use multi_gpu::GpuTaskResult;

pub use multi_gpu::LoadBalancingStrategy;

pub use multi_gpu::MultiGpuConfig;

pub use multi_gpu::MultiGpuConfigFactory;

pub use multi_gpu::MultiGpuManager;

pub use multi_gpu::MultiGpuStats;

pub use multi_gpu::MultiGpuTask;

pub use multi_gpu::TaskPriority;

pub use performance::GpuPerformanceStats;

pub use types::GpuExecutionConfig;

pub use kernels::*;

Modules

accelerator

Main GPU accelerator implementation

buffer

GPU memory buffer management

config

GPU configuration structures and enums

device

GPU device information and management

index

GPU-accelerated vector index implementations

index_builder

GPU-accelerated HNSW index construction

kernels

CUDA kernel implementations for various vector operations

load_balancer

GPU load balancing for distributing index-building work across multiple devices.

memory_pool

GPU memory pool management for efficient allocation and reuse

multi_gpu

Multi-GPU load balancing for distributed vector index operations

performance

GPU performance monitoring and statistics

runtime

CUDA runtime types and utilities

types

Common GPU types and structures