Native Acceleration

The native package provides automated building and loading of C++/CUDA native PyTorch extensions. There is no separate build step; importing an aggregator or attack that offers a native variant triggers compilation transparently.

For the design rationale behind import-time compilation — why the project trades a separate build system for zero-friction prototyping — see the explanation of /explanation/native-compilation.

Overview

Native Components

Component	Description	Environment Control
Module loader	Scans subdirectories, compiles sources, injects py_ modules	NATIVE_OPT, NATIVE_STD, NATIVE_QUIET
Dependency resolver	Reads .deps files and builds dependencies recursively	—
CUDA support	Detects .cu files and compiles with NVCC when available	—
Thread pool	Shared C++ threadpool dependency (so_threadpool) for parallel ops	—

Available Native Variants

When compilation succeeds, native variants are registered alongside their Python counterparts with a native- prefix:

Name	What is accelerated
native-krum	Pairwise distance computation and scoring.
native-median	Coordinate-wise median.
native-bulyan	Multi-Krum selection and median averaging.
native-brute	Combinatorial subset search.

Because the high-level API is identical, you can compare the exact same rule in Python and native versions simply by changing the registered name.

Internal Dependencies

Some modules under native/ are shared libraries (so_ prefix) that provide infrastructure for other native variants rather than exposing a public API:

• so_threadpool — C++ thread pool used by native-bulyan for parallel distance computations.

These libraries are built automatically when required by a dependent module.

API Reference

Native (i.e. C++/CUDA) implementations automated building and loading.