GoKawiilNew open-source tool rGPU lets PyTorch tensors run on a remote NVIDIA GPU
rGPU is a newly shown project that lets a PyTorch program send tensor operations over TCP to execute on a remote NVIDIA GPU while the application itself keeps running locally. It offers two integration routes: a simple PyTorch 'rgpu' device for programs that opt in, and a CUDA shim that intercepts existing Linux CUDA binaries using libcuda, CUDA Runtime, cuBLAS, cuBLASLt, and cuDNN. Setup involves installing via pip and using an SSH tunnel through the rgpu-run command to connect to the GPU host.
GoKawiil's interpretation of the reporting above, not reported fact.
By separating where code runs from where the GPU physically sits, rGPU could let developers use remote or shared GPU hardware without rewriting their PyTorch code or binaries, which may simplify cloud GPU access for smaller teams. The project's own documentation flags that neither protocol currently authenticates or encrypts traffic, meaning users must rely on SSH tunnels and firewall rules to avoid exposing GPU servers on the network. As an early 'Show HN' release, its compatibility and security model will likely need further hardening before broader production use.
- rGPU lets PyTorch tensors execute on a remote GPU via a new 'rgpu' device or a CUDA shim for existing binaries.
- The CUDA shim supports libcuda, CUDA Runtime, cuBLAS, cuBLASLt, and cuDNN, extending compatibility to unmodified CUDA programs.
- Neither the PyTorch device nor CUDA shim protocol is authenticated or encrypted, requiring SSH tunnels and firewall restrictions for safe use.
Source: github.com, 2026-10-07
Published there as: “Show HN: Rgpu – a PyTorch device whose tensors live on a remote GPU”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.