Joint Structural Pruning and Mixed-Precision Quantization for LLM Compression

In three linesNovel LLM compression framework jointly optimizing structural pruning and mixed-precision quantization. Minimizes global error propagation across the entire model rather than per-layer. At 1-3 bits, reduces WikiText perplexity by 21% vs SoTA weight-activation baselines and 59-85% vs weight-only methods.

Summary generated by Claude — human-verified