arXiv Machine Learning

When Local Variance Optimality Is Not Enough: RoPE-Aligned Q/K Rotations for Dynamic 4-Bit Quantisation

arXiv:2608. 13365v1 Announce Type: new Abstract: Rotation-based post-training quantisation commonly applies an orthogonal transform across an entire attention head to reduce outlier-induced error.

arXiv AI
Sep 4

HARP: Hadamard-Preconditioned Adaptive Rotation Processor for Extreme LLM Quantization

HARP (Hadamard‑Preconditioned Adaptive Rotation Processor) is a learnable, structured two‑sided orthogonal processor that replaces fixed randomized Hadamard transforms in post‑training quantization of large language models. By representing rotations as sparse butterfly‑like block‑orthogonal stages and supporting mixed‑radix schedules, HARP adapts the quantization basis to each layer and calibration distribution while maintaining full‑precision equivalence. Across 2–4‑bit settings on Llama models from 1B to 70B, HARP consistently improves perplexity, delivers the strongest zero‑shot gains at 2 bits, and preserves deployment efficiency—achieving 128 tokens per second on Llama 2 7B at 2 bits, roughly 90% of RHT throughput and over twice the speed of FP16.

By Artur Zagitov, Gleb Molodtsov, Aleksandr Beznosikov
arXiv Machine Learning
Aug 27

Transforms for LLM Quantization: The Great Inversion and Format Co-Design

The paper surveys the use of linear, function‑preserving transforms in 4‑bit large‑language‑model (LLM) quantization, formalizing the underlying principle as the "Great Inversion"—the trade‑off between energy concentration favored by allocation‑flexible coding and within‑group flattening favored by grouped shared‑scale quantization. It reviews 200 works, classifies 43 transform methods by structure, data‑awareness, construction approach, and runtime cost, and examines how they interact with GPTQ rounding. The study also explores how different number formats (FP4, MXFP4, NVFP4) influence the optimal transform choice and outlines open research problems. "whyItMatters":"The survey clarifies the conflicting objectives in transform‑based LLM quantization and provides a practical guide for selecting transforms based on deployment regime, thereby informing future research and deployment strategies."

By Ehsan Jokar
arXiv Machine Learning
Aug 10

{\Omega}-QVLA: Robust Quantization for Vision-Language-Action Models via Composite Rotation and Per-step Scaling

arXiv:2605. 28803v2 Announce Type: replace-cross Abstract: Vision-Language-Action (VLA) models unify perception, reasoning, and control within a single policy, yet their multi-billion-parameter backbones and diffusion-based action heads make on-device deployment prohibitively expensive.

By Xinyu Wang, Mingze Li, Sicheng Lyu, Dongxiu Liu, Kaicheng Yang, Ziyu Zhao, Yufei Cui, Xiao-Wen Chang, Peng Lu
arXiv Machine Learning
Jun 25

Why Do Accumulated Transformations Extrapolate?

arXiv:2606. 24975v1 Announce Type: new Abstract: PaTH Attention showed that replacing RoPE's position-indexed rotations with accumulated data-dependent Householder reflections yields strong length extrapolation, though performance degrades at extreme context lengths.

By Mahesh Godavarti
arXiv AI
Jul 7

HiFA4: Training-Free 4-bit FlashAttention on Ascend HIF4 NPUs for LLM Inference

arXiv:2607. 04302v1 Announce Type: cross Abstract: We present HiFA4, a post-training operator-level design that executes both QK^T and PV in FlashAttention as 4-bit HIF4 Cube GEMMs for LLM inference on Ascend NPUs, while maintaining the online softmax state in FP16.

By Hui Dong, Yanzhao Li, Jie Gao, Chunlu Li, Zhiyuan Zhang, Yupeng Sun, Zhenyuan Chen, Zhiqiang Zou
arXiv Machine Learning
Aug 19

CORAM: Coherent Orthogonal Rotation for Model Merging

CORAM (Coherent Orthogonal Rotation for Model Merging) is a new method for combining fine‑tuned models without joint training or access to original data. It partitions each target weight matrix into row slices, represents each expert slice with its singular value decomposition in the base‑model SVD frame, and merges the task‑specific factors on their corresponding manifolds. The approach includes an amplification coefficient to counteract manifold averaging contraction, spread slicing to balance highly updated rows, and a residual pathway for non‑target layers, achieving improvements over existing orthogonal merging techniques across multiple model families and scales.

By Xinyi Sui, Ziran Liu, Nam Ling, Wei Wang, Wei Jiang