Graphzero

Zero-copy C++ graph engine to train PyTorch GNNs with 0 RAM.

Share

GraphZero fixes PyTorch Geometric OOM crashes by memory-mapping massive graph datasets directly from your SSD. Built with C++20 and nanobind, it hands raw pointers to PyTorch as zero-copy NumPy arrays. Train 50GB models on consumer hardware by letting the OS handle page faults while the GPU focuses entirely on the math.