This project provides a way to run PyTorch operations and keep tensors on a remote NVIDIA GPU while the application runs on a local client machine (for example, a Mac without CUDA). It exposes two integration paths: a PyTorch device that transparently places tensors on a remote GPU for programs that opt in, and a CUDA shim that implements libcuda, the CUDA runtime and key libraries (cuBLAS, cuDNN, etc.) so existing Linux CUDA binaries can use a remote GPU without recompilation. The simpler PyTorch device is intended for new or adapted PyTorch workloads; the CUDA shim targets compatibility with existing CUDA programs. A command-line helper opens an SSH tunnel and configures the connection; a minimal example shows creating a tensor on the remote device and performing operations locally while execution happens remotely.
The repository includes client and server implementations, shared protocol metadata, a Python package providing the device and launcher, tests, build scripts for the C++ client and CUDA shim, and product documentation with quickstarts, examples (including training and a nanoGPT example), performance notes and deployment guidance. Neither protocol authenticates or encrypts by default, so the recommended deployment uses SSH tunnels and firewall rules (the CUDA server listens on port 9713 by default) to restrict access. Source is Apache-2.0 licensed and includes generated C++ artifacts and development tooling for testing and building.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
NanoMuse is an open-source AI agent designed to work across multiple devices, including phones, tablets, and computers. It remembers user preferences, stays active when apps are closed, and interacts through a personalized interface. (github.com)
OpenWAM is an open framework that combines video prediction and robot control using a shared causal video backbone. It supports various interaction models and improves robot task success rates through causal adaptation and counterfactual supervision. (openwam.stanford.edu)
Building Rome from a Single Image creates detailed 3D scene reconstructions using a novel pipeline that generates meshes from just one photo. The method employs adaptive scene chunks, explicit 2D-3D conditioning, and autoregressive generation to produce consistent, high-quality models. (build-rome.github.io)
OpenAI's Decisions API can evaluate GitHub pull requests by analyzing diffs against specific review questions, such as compatibility or migration needs. This helps organize review processes by providing clear answers to targeted concerns. (vercel.com)
Strands Decider 2B is an open-source decision model designed for building AI agents. It supports various tools, integrations, and deployment options for creating and managing intelligent systems. (strandsagents.com)
UniEvo-VL is a framework that enables multimodal models to improve themselves by using their own feedback during testing. It enhances image generation performance without external supervision, showing significant gains on benchmark tests. (arxiv.org)
Today's best Hacker News stories, summarized and screenshotted, one email a day.