Why Gated DeltaNet Survives 4-Bit Quantization: NVFP4 W4A4 for the Recurrent Half of a Hybrid 27B LLM

📄 arXiv: 2609.04098v1 📥 PDF

作者: Sergii Kozyrev, Davyd Maiboroda

分类: cs.AI

发布日期: 2026-09-03

备注: 14 pages, 2 figures, 6 tables. Quantized checkpoint: https://huggingface.co/minima-ai/mnma_qwen3.8_27b_nvfp4

🔗 代码/项目: HUGGINGFACE


💡 一句话要点

提出NVFP4 W4A4以解决混合27B LLM的4位量化问题

🎯 匹配领域: 支柱二:RL算法与架构 (RL & Architecture) 支柱八:物理动画 (Physics-based Animation)

关键词: 混合LLM 量化方法 Gated DeltaNet 自然语言处理 模型优化 深度学习 性能提升

📋 核心要点

  1. 现有的4位量化方法未能有效处理Gated DeltaNet模块,导致性能损失。
  2. 提出NVFP4 W4A4方法,通过量化所有线性层来提高模型的性能和效率。
  3. 实验结果表明,Minima在多个基准测试中表现出色,且在模型大小和速度上具有优势。

📝 摘要(中文)

混合LLM结合了softmax注意力与线性注意力层,如Gated DeltaNet(GDN),其递归状态以固定大小总结上下文。早期的4位量化尝试未能有效量化GDN模块,尤其是其衰减和写入强度门。本文通过构建Minima:NVFP4 W4A4,在所有496个线性层上进行测试,结果显示在多个基准测试中,Minima的表现与BF16相当,同时在模型大小和预填充速度上具有优势。通过四个部分的机制研究,本文解释了为何GDN在量化过程中表现良好,并提出了一种实用的量化方案。

🔬 方法详解

问题定义:本文旨在解决混合LLM中Gated DeltaNet模块在4位量化时的性能损失问题。现有方法在量化过程中未能有效处理GDN模块,尤其是其衰减和写入强度门,导致错误积累。

核心思路:论文提出NVFP4 W4A4方法,通过在所有496个线性层上进行量化,旨在提高模型的整体性能和效率。该设计基于对递归状态的噪声处理能力的理解。

技术框架:整体架构包括对GDN模块的量化处理,采用16元素块缩放来局部化残差流的极端异常值,并通过四个部分的机制研究来解释其有效性。

关键创新:最重要的技术创新在于通过量化所有线性层而非仅限于GDN模块,从而实现了更均衡的激活误差,显著提升了模型的鲁棒性和性能。

关键设计:在参数设置上,采用softplus/exponential和sigmoid参数化方法,压缩了GEMM误差,并通过delta-rule递归保持注入噪声的平稳性,确保在长上下文中性能稳定。

🖼️ 关键图片

img_0
img_1
img_2

📊 实验亮点

实验结果显示,Minima在多个基准测试中与BF16的表现相当,5任务平均误差为-0.52,同时在模型大小(17.5 GiB)和预填充速度上提升了14-19%。32K困惑度差距随着位置的变化而缩小,表明量化效果的稳定性。

🎯 应用场景

该研究的潜在应用领域包括自然语言处理、对话系统和其他需要高效模型的AI应用。通过优化量化过程,能够在资源受限的环境中部署大型语言模型,提升实际应用的可行性和效率。

📄 摘要(原文)

Hybrid LLMs pair softmax attention with linear-attention layers such as Gated DeltaNet (GDN), whose recurrent state summarizes the context in fixed size. Early community 4-bit quantizations of Qwen3.8-27B (48 GDN layers, 16 attention layers) left the GDN block in 8- or 16-bit precision -- especially its decay and write-strength gates -- on the intuition that errors in a recurrence accumulate over long contexts. We test that intuition by building Minima: NVFP4 W4A4 on all 496 linear layers, GDN included. Across perplexity at 4K/32K, MMLU-Pro, GSM8K, AIME'25, GPQA-Diamond, LiveCodeBench, and RULER retrieval to 64K, Minima matches BF16 within seed noise (5-task average -0.52) while being the smallest (17.5 GiB) and fastest-prefill (+14-19%) recipe we compare, and its 32K perplexity gap shrinks with position. A four-part mechanism study explains why: (i) NVFP4's 16-element block scaling localizes the residual stream's extreme outliers, equalizing activation error across layer roles; (ii) the supposedly fragile gate projections are the least sensitive -- softplus/exponential and sigmoid parameterizations compress ~11% GEMM error to ~2% output error; (iii) the delta-rule recurrence holds injected noise at a flat plateau over 32K tokens and forgets a state impulse within hundreds of steps, because each write overwrites the state along the current key direction; (iv) the per-token quantization cost washes out with context instead of compounding. We also repair a global-scale mismatch that arises when per-module-calibrated NVFP4 checkpoints are served by kernels that fuse those modules into one GEMM, and show calibrated FP8 KV-cache scales are performance-free. The result: a practical recipe -- quantize everything, ship KV scales -- and a mechanistic account of why the recurrent half of a hybrid LLM is the easy half to quantize. Checkpoint: https://huggingface.co/minima-ai/mnma_qwen3.8_27b_nvfp4