[P] I got tired of PyTorch Geometric OOMing my laptop, so I wrote a C++ zero-copy graph engine to bypass RAM entirely.

A developer has open-sourced GraphZero v0.2, a custom C++ data engine designed to eliminate the massive memory bottlenecks plaguing Graph Neural Network (GNN) training. Standard libraries like PyTorch Geometric attempt to load entire datasets—edge lists and feature matrices—into system RAM, instantly crashing with 24GB+ allocations on large datasets like Papers100M. GraphZero solves this by bypassing RAM entirely. It compiles raw CSV data into two highly optimized binary formats (.gl for graph topology, .gd for node features) and uses the POSIX `mmap` system call to memory-map these massive files directly from the SSD.

During training, the C++ engine, bound to Python via nanobind, hands raw memory pointers directly to PyTorch as zero-copy NumPy arrays. When PyTorch indexes a batch, it triggers an OS page fault, and the operating system fetches only the required 4KB blocks from the NVMe drive on-demand. To keep the pipeline efficient, the engine uses OpenMP to multi-thread operations like neighbor sampling (`batch_random_fanout`) and releases the Python Global Interpreter Lock (GIL), allowing full parallelization of disk I/O, CPU sampling, and GPU computations. The result is the ability to train on a 50GB dataset while Python allocates zero bytes of RAM for the dataset itself, a breakthrough for researchers with limited hardware.