SparseLab fits 1B models in 400MB RAM

SparseLab solves the "fake sparsity" problem by actually freeing up memory, making it a critical tool for training large models on edge devices and consumer laptops.

• –Implements Padded-CSR format for O(1) topology mutation without full matrix reallocations.
• –Ships with SET and RigL algorithms to evolve network connections dynamically during the training process.
• –Native support for Apple Silicon NEON and Linux OpenMP enables high-memory research on commodity hardware.
• –The 4x speed penalty is a significant bottleneck, but the library targets developers constrained by VRAM/RAM rather than compute cycles.
• –Drop-in nn.Linear replacement makes it trivial to port existing dense architectures to sparse-native training.