Lupine — удалённый GPU как локальный

★ 2.4K

lupine is an open-source GPU-over-IP bridge that attaches a remote machine's GPU to a CPU-only server, making the remote card appear local to any process without modifying code or models. Reach for it when a powerful workstation GPU sits idle overnight but inference needs to run on a headless CPU server; when GPUs are spread across several machines and you want a shared pool without moving data or environments; or when buying a dedicated GPU server isn't justified for a small team. Deployed via Docker: run the server image on the GPU machine (`--gpus all`, port 14833) and point the client container at it via `LUPINE_SERVER`. Python projects integrate transparently with `pip install "lupine[auto]"`—no explicit LUPINE API calls needed. Supports Linux, macOS, and Windows on amd64 and arm64; macOS gets CUDA/NVML shims loadable via `ctypes`, but PyTorch requires a CUDA-enabled Linux or Windows build. Server images are published to GHCR under `cuda-<version>-ubuntu<version>` tags. The server handles graceful shutdown via SIGTERM with in-flight CUDA call draining, and supports an optional checkpoint provider ABI. The direction is forwarding a remote GPU to a local CPU process—not load-bal