rgpu-run

rgpu-run --host DEST [--ssh-port N] [-i KEY] [--remote-port N] [--timeout SECS] [--] command ...rgpu-run --server HOST:PORT [--] command ...

rgpu-run is the launcher from the rgpu Python package: a PyTorch device whose tensors live on a remote NVIDIA GPU. The application stays on the client; GPU work runs on the host. The script still has to `import rgpu` and select `device="rgpu"`; this command only opens the connection plumbing.Install with pip install rgpu (Python 3.10+, torch>=2.14). Console scripts are rgpu-run (`rgpurun:main`) and **rgpu-opserver** (`rgpu.server.main_:main`). Package version 0.1.1. The GPU host needs its own PyTorch of the same major.minor as the client.With --host, rgpu-run binds a free local port, starts `ssh -N -L local:127.0.0.1:remote`, waits until something accepts on the local end, runs the command with RGPU_OPSERVER pointing at that address, then terminates the tunnel (SIGTERM, then SIGKILL after 5s). Both ends of the forward are 127.0.0.1 on purpose: rgpu-opserver has no authentication.A second path, a CUDA shim (`libcuda` and related libraries), covers existing Linux CUDA binaries; that is built with `./scripts/build_client.sh` in the repository, not by this launcher.

--host DEST

SSH destination of the GPU host, for example `user@gpuhost`. Mutually exclusive with --server. Opens `ssh -N` with `ExitOnForwardFailure=yes` and forwards a free local port to `127.0.0.1:remote-port` on the host

Address of an already-reachable rgpu-opserver. No tunnel is opened. Mutually exclusive with --host

SSH port on the GPU host (passed as `ssh -p`)

SSH private key (passed as `ssh -i`)

Port rgpu-opserver listens on. Default 9720

Seconds to wait for the tunnel to accept connections. Default 30

Command to run after the tunnel is up (or immediately with --server). A leading `--` is stripped. The child inherits the environment with RGPU_OPSERVER set to `127.0.0.1:local-port` (tunnel) or the --server address. Missing command: argparse error. Missing executable: exit 127. KeyboardInterrupt: exit 130

Neither protocol authenticates or encrypts connections. Keep rgpu-opserver on its default localhost bind and reach it through SSH. The CUDA server listens on all IPv4 interfaces: restrict port 9713 with host or cloud firewall rules before starting it. The server PyTorch version must match the client's major.minor.

rGPU is an Apache-2.0 project by Yan Michalevsky. The Python package is rgpu. Documentation lives at rgpu.dev.

python(1), pip(1), ssh(1), pytorch(1)

Source code · Homepage · Documentation