E-PMQ: Expert-Guided Post-Merge Quantization with Merged-Weight Anchoring

In three linesE-PMQ introduces expert-guided post-merge quantization to efficiently deploy multi-expert merged models at low precision. Using source expert weights as calibration targets and merged-weight anchoring, E-PMQ improves 4-bit GPTQ from 65.0% to 73.6% on CLIP-ViT-B/32 (8-task merging) and from 34.8% to 76.7% on CLIP-ViT-L/14 (20-task merging).

Summary generated by Claude — human-verified