Depth Registers Unlock W4A4 on SwiGLU: A Reader/Generator Decomposition

• URL: http://arxiv.org/abs/2604.18128v1
• Date: Mon, 20 Apr 2026 11:47:11 GMT
• Title: Depth Registers Unlock W4A4 on SwiGLU: A Reader/Generator Decomposition
• Abstract summary: We study post-training W4A4 quantization in a SwiGLU decoder-only language model trained on 5B tokens of FineWeb-Edu.<n>Naive round-to-nearest W4A4 collapses validation perplexity from FP16 23.6 to 1727.<n>We present DR+sink as a training-time probe rather than a deployment proposal.
• Score: 2.6382975801439836
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: We study post-training W4A4 quantization in a controlled 300M-parameter SwiGLU decoder-only language model trained on 5B tokens of FineWeb-Edu, and ask which input-activation sites dominate the error. Naive round-to-nearest W4A4 collapses validation perplexity from FP16 23.6 to 1727. A simple residual-axis training-time intervention -- Depth Registers with a register-magnitude hinge loss (DR+sink) -- reduces this to 119 (about 14x) at matched FP16 PPL and matched zero-shot capacity, and composes with SmoothQuant to 39.9 PPL. The residual ~2 PPL gap to FP16 is the diagnostic core. We decompose W4A4 damage by input-activation site: the five trainable linears in a SwiGLU block split into residual-axis readers (qkv, w1, w3) and block-internal generators (o_proj, w2). Elementary norm arguments show residual-axis magnitude control bounds readers tightly but leaves w2's bilinear input bounded only by the trivial product of factor bounds; empirically, DR+sink collapses reader kurtosis while leaving generators essentially unchanged, and the reader-rescued W4A4 residue is flat at ~0.28 nats across three matched checkpoints with Delta-remove(w2) dominating. We present DR+sink as a training-time probe rather than a deployment proposal: a post-hoc alternative (Per-Linear QuaRot) nearly matches it on the reader axis. Full QuaRot -- adding online per-head value Hadamard plus online w2-input rotation -- does not close the gap either, directly testing the prediction that orthogonal rotation cannot bound the bilinear SwiGLU tail. Claims are specific to our 300M, 5B-token, single-seed setting, and our experiments do not isolate the partition from the hinge.

Related papers

• VC-Attention: Value Smoothing and Softmax Casting for Low-bit Attention [47.30761996669105]
  We propose VC-Attention, a training-free low-bit attention framework.<n>It addresses both by smoothing value tokens by a probability Cast. V-Smooth reorders value tokens by lightweight clustering online, so the tokens in a fused hardware block quantize well together.<n>It maps log-domain scores directly to E4M3 probability codes with one fused multiply-add, eliminating the FP32 exponential and the format conversion.
  arXiv  Detail & Related papers  (2026-09-14T16:17:35Z)
• Why Gated DeltaNet Survives 4-Bit Quantization: NVFP4 W4A4 for the Recurrent Half of a Hybrid 27B LLM [0.0]
  Minima: NVFP4 W4A4 on all 496 linear layers, GDN included.<n>A four-part mechanism study explains why NVFP4's 16-element block scaling localizes the residual stream's extreme outliers.<n>A mechanistic account of why the recurrent half of a hybrid LLM is the easy half to quantize.
  arXiv  Detail & Related papers  (2026-09-03T17:04:26Z)
• HiFloat4 Format for End-To-End Reinforcement Learning Post-Training of Large Language Models [9.149747042283197]
  We present the first end-to-end FP4 RL post-training, in which both the rollout and training policies operate at 4-bit precision.<n>A lightweight correction, Rollout-ResQ, recovers most of the precision lost to outlier-driven underflow without inflating the rollout's compute footprint.
  arXiv  Detail & Related papers  (2026-07-29T06:28:26Z)
• QUADS: Stabilizing NVFP4 Reinforcement Learning for MoE via QUantization-error Alignment across Dual Sides [26.939321070753607]
  Rollout generation is a major bottleneck in Reinforcement Learning for Mixture-of-Experts (MoE) Large Language Models.<n>We propose QUantization-error Alignment across Dual Sides (QUADS) to stabilize NVFP4 RL for MoE.<n>QUADS achieves BF16-level accuracy, improves average pass@1 by 21.49 points over naive NVFP4 RL, and delivers 16% higher rollout throughput than FP8.
  arXiv  Detail & Related papers  (2026-07-17T10:21:27Z)
• Right Makes Might: Aligning Verified Hidden States Empowers RL Reasoning [55.264863369127774]
  Current methods reduce each correct rollout to a single reward bit, ignoring the geometric structure shared among their hidden states.<n>We propose Hidden-Align, an auxiliary loss function that aligns the last-layer hidden states of correct rollouts at the anchor token during RL training, with zero overhead in both training and inference.<n>On eight mathematical reasoning benchmarks, Hidden-Align improves average pass@1 over the DAPO baseline by 3.8, 6.2, and 5.4 percentage points on Qwen3-1.7B, 4B, and 14B respectively, with consistent pass@k gains across all three scales.
  arXiv  Detail & Related papers  (2026-06-02T06:51:15Z)
• Influence-Inspired Spectral Rotations for Extreme Low-Bit LLM Quantization [0.0]
  We apply the influence of a companion theory paper to extreme low-bit weight-only quantization.<n>We show that BBTspectral reduces wikitext-2 perplexity by 58% relative to vanilla autoround at W2A16.
  arXiv  Detail & Related papers  (2026-05-24T18:05:37Z)
• Amplifying, Not Learning: Fine-Tuned AI Text Detectors Amplify a Pretrained Direction [51.56484100374058]
  Text detectors amplify a pretrained typicality axis; they do not construct an AI-vs-human boundary.<n>On raw encoders before any task supervision, projecting onto centroid(AI)-centroid(HC3) achieves NYT-vs-HC3 AUROC 0.806/0.944/0.834 across three architectures.<n>On RoBERTa-base, raw projection exceeds fine-tuning; on RoBERTa-base, full fine-tuning reduces discrimination below raw on both fluent-formal populations tested.
  arXiv  Detail & Related papers  (2026-05-20T19:08:38Z)
• DuQuant++: Fine-grained Rotation Enhances Microscaling FP4 Quantization [47.19478866645546]
  We propose DuQuant++, which adapts the outlier-aware fine-grained rotation of DuQuant to the MXFP4 format.<n>Experiments on the LLaMA-3 family under MXFP4 W4A4 quantization show that DuQuant++ consistently achieves state-of-the-art performance.
  arXiv  Detail & Related papers  (2026-04-20T04:27:28Z)
• When Flat Minima Fail: Characterizing INT4 Quantization Collapse After FP32 Convergence [0.0]
  Post-training quantization assumes that a well-converged model is a quantization-ready model.<n>We show this assumption fails in a structured, measurable, and previously uncharacterized way.
  arXiv  Detail & Related papers  (2026-04-16T15:46:36Z)
• Auxiliary Finite-Difference Residual-Gradient Regularization for PINNs [0.0]
  We study a hybrid design in which the governing PDE residual remains automatic-differentiation based, while finite differences appear only in a weak auxiliary term.<n>The FD term regularizes the residual field without replacing the PDE residual itself.<n>Across seeds 0-5 and 100k epochs, the most reliable tested configuration is a fixed shell weight of 5e-4 under the Kourkoutas-beta regime.
  arXiv  Detail & Related papers  (2026-04-15T23:09:19Z)
• Dissecting Outlier Dynamics in LLM NVFP4 Pretraining [46.10969678564592]
  This study conducts a longitudinal analysis of outlier dynamics across architecture during NVFP4 pretraining.<n>We find that, compared with Softmax Attention (SA), Linear Attention (LA) reduces per-tensor heavy tails but still exhibits persistent block-level spikes under block quantization.<n>We then develop CHON, an NVFP4 training recipe integrating with post-QK operation protection.
  arXiv  Detail & Related papers  (2026-02-02T12:50:27Z)
• Ensemble Threshold Calibration for Stable Sensitivity Control [0.0]
  We present an end-to-end framework that achieves exact recall with sub-percent variance over tens of millions of geometry pairs.<n>Our approach consistently hits a recall target within a small error, decreases redundant verifications relative to other calibrations, and runs end-to-end on a single TPU v3 core.
  arXiv  Detail & Related papers  (2025-10-02T15:22:28Z)
• SVDQuant: Absorbing Outliers by Low-Rank Components for 4-Bit Diffusion Models [61.474101404805545]
  Diffusion models can generate high-quality images, but as they scale, rising memory demands and higher latency pose deployment challenges.<n>We propose SVDQuant, a new 4-bit quantization paradigm to overcome this limitation.<n>We reduce the memory usage for the 12B FLUX.1 models by 3.5$times$, achieving 3.0$times$ speedup over the 4-bit weight-only quantization (W4A16) baseline.
  arXiv  Detail & Related papers  (2024-11-07T18:59:58Z)
• SpinQuant: LLM quantization with learned rotations [49.07335692298487]
  Post-training quantization (PTQ) techniques applied to weights, activations, and the KV cache greatly reduce memory usage, latency, and power consumption of Large Language Models (LLMs)<n>We identify a collection of applicable rotation parameterizations that lead to identical outputs in full-precision Transformer architectures while enhancing quantization accuracy.<n>We propose SpinQuant, a novel approach that incorporates learned rotation matrices for optimal quantized network accuracy.
  arXiv  Detail & Related papers  (2024-05-26T02:15:49Z)
• QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs [72.26197676852958]
  We introduce QuaRot, a new Quantization scheme based on Rotations. QuaRot quantizes end-to-end, including all weights, activations, and KV cache in 4 bits. Our 4-bit quantized LLaMa2-70B model has losses of at most 0.47 WikiText-2 perplexity and retains 99% of the zero-shot performance.
  arXiv  Detail & Related papers  (2024-03-30T19:20:06Z)
• Calibration of Drive Non-Linearity for Arbitrary-Angle Single-Qubit Gates Using Error Amplification [43.97138136532209]
  Non-linearity of qubit drive line components imposes a limit on the fidelity of single-qubit gates. We demonstrate arbitrary-angle single-qubit gates with coherence-limited errors of $2times 10-4$ and leakage below $6times 10-5$.
  arXiv  Detail & Related papers  (2022-12-02T10:34:43Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.