70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float Paper • 2504.11651 • Published Apr 15, 2025 • 31
