NVlabs/vibetensor

Implements a complete deep learning runtime stack spanning Python/Node.js bindings through a C++20 core with native CUDA support, including a schema-lite dispatcher, reverse-mode autograd, stream-ordered memory allocator, and CUDA graph capture. Provides PyTorch-compatible eager execution with DLPack zero-copy interoperability, dynamic operator plugins via stable C ABI, and experimental multi-GPU fabric tensors with topology awareness. The entire ~60k-line C++/CUDA codebase was generated and validated by LLM agents using build tests and differential checks rather than manual code review.

578 stars.