SoLA: Leveraging Soft Activation Sparsity and Low-Rank Decomposition for Large Language Model Compression

Implement a training-free compression engine that uses soft activation sparsity to shrink model size for consumer GPUs. This provides a drop-in replacement for standard quantization methods.

Suggested repo: SoLA-Engine

"Slim your models without the training hassle or hardware-specific requirements."

Estimated effort: 60h