Rotation Is Token-Blind: A Structural Criterion for Post-Training Quantization of Feed-Forward Geometry Transformers
Read the original on arXiv Computer Vision →The paper introduces QMapAnything, a post‑training quantization framework tailored for feed‑forward geometry transformers that use Transformers to predict camera poses, depth, and structure in a single pass. It addresses the challenge that heterogeneous token types share a single activation quantization grid, which can cause high‑leverage tokens to dominate outputs. By treating rotation and smoothing as token‑blind reparameterizations and defining a statistic γ to gauge when differentiated bias correction is needed, the method achieves W4A8 quantization that closely matches full‑precision performance while W8A8 reduces peak memory by 2.2–3.1× and speeds up inference by up to 3.1× on real hardware.
Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Computer Vision.