cuVS build and RunPod workflow

Build cuVS locally, run tests and a smoke benchmark on RunPod, and download results. No local GPU is required. The cuVS checkout uses a dedicated cpp/build-runpod directory.

Setup

• macOS / ARM: Python 3, Make, Docker, and the builder image. Make commands automatically use Linux amd64 Docker.
• Linux x86_64: Python 3, Conda, Make, SSH, and GNU rsync.
• RunPod: Linux x86_64 with an NVIDIA GPU, driver, Python 3, and rsync.

cp .env.example .env
cp runpod.ssh.example runpod.ssh

Set CUVS_SOURCE_DIR in .env and the pod address, port, and identity in runpod.ssh. Unlock passphrase-protected keys with ssh-add, then run:

make setup
make cycle

Development

After editing cuVS, run make cycle: build → package → upload → test → benchmark → download results. Reports are saved in results/.

Command	Purpose
make build	Compile selected targets.
make publish	Build, package, and upload.
make test / make benchmark	Run the last published release.
make fetch	Download reports.
make inventory	Refresh hardware information after changing pods.
make check	Check the workflow without SSH or a GPU.
make help	Show all commands.

Edit config/build.json for build and test settings. The default runs quick smoke tests; use make cycle CONFIG=config/build-all-native.json for the full profile. See Docker options and logs for detached builds and compiler job limits.

Library cache

Attach a network volume at /workspace. Libraries are reused from /workspace/cuvs-library-cache; only missing or changed content is uploaded. Keep this cache when removing old releases, and attach the same volume to replacement pods. See cache configuration and maintenance.