Show HN: Rgpu – a PyTorch device whose tensors live on a remote GPU

Category: infrastructure

Tags: remote-gpu, pytorch, cuda, distributed-computing, gpu-offloading

Score: 7.0/10 (Innovation: 7, Technical: 8, Documentation: 7, Utility: 6)

rGPU allows PyTorch and CUDA programs to offload GPU computation to a remote NVIDIA machine over the network, either via a custom PyTorch device or a CUDA shim that intercepts libcuda/cuBLAS/cuDNN calls. It is technically ambitious, providing a transparent remote GPU experience, but requires careful setup and has security limitations.

Target audience: machine learning engineers, researchers, devops

Repository: https://github.com/ymcrcat/rgpu · C++ · Apache-2.0 · 79 stars

View on Hacker News