@sweet-search/native-linux-arm64-gnu-cuda

NVIDIA CUDA-enabled native binaries for sweet-search on Linux arm64 (glibc) — Jetson Orin, Grace Hopper, and arm64 SBSA server GPUs.

This package is a platform-scoped optionalDependency of the main sweet-search package. npm installs it automatically on matching hosts:

• os === 'linux'
• cpu === 'arm64'
• libc === 'glibc' (Ubuntu/Debian/RHEL on arm64 — not Alpine/musl)

On non-matching platforms, npm silently skips this package.

What's inside

• sweet-search-native.node — napi-rs addon built with the cuda,flash-attn Cargo features for aarch64-unknown-linux-gnu. Embedding + late-interaction inference dispatch to candle-cuda when libcuda.so.1 is present at runtime; otherwise the addon fails to load and sweet-search falls back to the CPU-only @sweet-search/native-linux-arm64-gnu variant.
• sweet-search — Rust CLI binary (identical to the non-CUDA Linux arm64 package).

Runtime requirements

• Linux arm64 (aarch64) with glibc ≥ 2.31
• NVIDIA driver providing libcuda.so.1
• Compute capability ≥ 7.0 (Jetson Orin is SM 8.7, Grace Hopper is SM 9.0; older Jetson Xavier at SM 7.2 also qualifies but is EOL)
• CUDA Toolkit 12.x runtime components; the binary is linked against CUDA 12.2 at build time

Flash-attention kernels are compiled for CUDA_COMPUTE_CAP=87 (Jetson Orin) with forward compatibility to SM 9.0 (Grace Hopper). On SM 7.x hardware the flash-attn path is skipped at runtime and candle's naive attention is used instead — the binary supports both paths.

Install

Automatic via the main sweet-search package:

npm install sweet-search

On a Linux arm64 host with a working NVIDIA driver, npm pulls this CUDA-enabled addon and core/infrastructure/native-resolver.js prefers it over the plain @sweet-search/native-linux-arm64-gnu variant.

Detecting whether CUDA is armed

Runtime detection is authoritative. Run sweet-search init — its report prints either NVIDIA GPU: <name> ... candle-cuda armed or a warning that the standard CPU-only package should be installed. .sweet-search/config.json records the decision under runtime.hardware. See docs/INIT_STRATEGY.md → "CUDA Backend" for the full contract.

Troubleshooting

Force-disable CUDA without uninstalling:

export SWEET_SEARCH_CUDA=0    # or pass --skip-cuda to `sweet-search init`

Parity between the CUDA path and the CPU reference is validated pre-release with node scripts/parity-cuda.js on a real Jetson Orin or Grace host.