Research Papers 论文研究 4h ago Updated 1h ago 更新于 1小时前 46

Signed Symmetric Quantization for Few-Bit Integers 用于少量整数的有符号对称量化

Signed Symmetric Quantization introduces a lightweight sign-selection rule that places the extra representable negative value on the dominant outlier tail, reducing clipping errors in few-bit precision. The method achieves conditional bound-optimality on L2 quantization error, holding true for 88-99% of weight groups in pre-trained LLMs at low bit widths. Negating the scale of a standard symmetric quantizer is analytically equivalent to a unit zero-point shift, offering a performance boost witho 提出有符号对称量化(Signed Symmetric Quantization),通过选择scale的符号将多出的负值表示范围分配给主导异常值尾部,解决低位宽量化中正异常值被截断的问题。 理论证明该方法在$\ell_2$量化误差上具有条件最优性,且在Qwen3、Llama3等主流大模型的低比特权重组中,该条件在88-99%的情况下成立。 实验表明,该方法在不增加推理成本的前提下,相比标准无符号对称量化显著降低了困惑度并提升了下游少样本准确率,同时保持了比非对称量化更高的吞吐量和更低的内存占用。

60
Hot 热度
75
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Signed Symmetric Quantization introduces a lightweight sign-selection rule that places the extra representable negative value on the dominant outlier tail, reducing clipping errors in few-bit precision.
  • The method achieves conditional bound-optimality on L2 quantization error, holding true for 88-99% of weight groups in pre-trained LLMs at low bit widths.
  • Negating the scale of a standard symmetric quantizer is analytically equivalent to a unit zero-point shift, offering a performance boost without the runtime penalties of asymmetric quantization.
  • Empirical validation on Qwen3, Qwen3.5, and Llama3 families demonstrates improved perplexity and few-shot accuracy compared to standard unsigned symmetric quantizers.

Why It Matters

This research provides a critical optimization for deploying large language models on resource-constrained hardware by bridging the gap between the speed of symmetric quantization and the accuracy of asymmetric methods. By eliminating the need for complex zero-point calculations during inference, it enables higher throughput and lower memory usage, which is essential for scalable AI deployment.

Technical Details

  • Problem Identification: Standard symmetric quantizers fix scales to be strictly positive, assigning the extra signed integer value to the negative tail. This causes clipping of positive outliers, leading to significant quantization errors at few-bit precisions.
  • Proposed Solution: Signed Symmetric Quantization utilizes a "signed absmax grid" where the scale sign is determined by a principled, lightweight rule to align the extra representable value with the dominant outlier direction.
  • Theoretical Analysis: The authors prove that this approach is conditionally bound-optimal for L2 error. They also demonstrate the mathematical equivalence between negating the scale in symmetric quantization and applying a unit zero-point shift in the signed integer alphabet.
  • Performance Benchmarks: On AMD EPYC "Turin" CPUs, 4-bit symmetric formats showed up to 9% less memory usage and 2.45x higher throughput than asymmetric counterparts. The proposed method maintains this efficiency while improving model accuracy.
  • Empirical Validation: Tested on Qwen3, Qwen3.5, and Llama3 models, showing consistent improvements in perplexity and downstream few-shot tasks without increasing inference latency.

Industry Insight

  • Hardware Efficiency: AI engineers should prioritize signed symmetric quantization for edge devices and high-throughput servers where memory bandwidth and compute efficiency are bottlenecks, as it avoids the overhead of asymmetric operations.
  • Model Compression Strategy: When moving to sub-8-bit precision, simply flipping the scale sign based on outlier distribution can yield immediate accuracy gains without retraining, serving as a drop-in replacement for standard symmetric quantization.
  • Future Standardization: As models grow larger, the cost of asymmetric quantization becomes prohibitive. This technique may become the default standard for efficient inference engines, necessitating updates to existing quantization libraries and compilers.

TL;DR

  • 提出有符号对称量化(Signed Symmetric Quantization),通过选择scale的符号将多出的负值表示范围分配给主导异常值尾部,解决低位宽量化中正异常值被截断的问题。
  • 理论证明该方法在$\ell_2$量化误差上具有条件最优性,且在Qwen3、Llama3等主流大模型的低比特权重组中,该条件在88-99%的情况下成立。
  • 实验表明,该方法在不增加推理成本的前提下,相比标准无符号对称量化显著降低了困惑度并提升了下游少样本准确率,同时保持了比非对称量化更高的吞吐量和更低的内存占用。

为什么值得看

本文针对大模型低比特量化中的关键痛点——正负不对称导致的精度损失,提供了一种无需额外计算开销的高效解决方案。对于致力于优化模型部署效率、平衡性能与速度的AI工程师而言,这一技术为在资源受限环境下提升模型表现提供了重要的理论依据和实践路径。

技术解析

  • 问题定义:标准对称量化强制scale为正,导致有符号整数中多出的一个负值表示范围被浪费在负尾,而正方向的异常值可能因超出范围被截断,造成显著的量化误差。
  • 核心方案:引入“signed absmax”网格,通过轻量级的符号选择规则,动态决定scale的正负,从而将额外的表示能力灵活分配给数据分布中异常值更密集的一侧,同时保持零点为零以维持对称量化的硬件友好性。
  • 理论分析:证明了有符号对称量化在$\ell_2$误差下的条件最优性,并通过数学推导指出,否定标准对称量化器的scale等价于在有符号整数字母表上进行单位零点偏移。
  • 实证验证:在AMD EPYC "Turin" CPU上,4位对称量化相比非对称量化节省高达9%内存且吞吐量高2.45倍;在Qwen3/Llama3系列模型上验证了其在保持推理速度优势的同时提升精度。

行业启示

  • 量化策略优化:在追求极致压缩率(如4bit及以下)时,简单的对称量化可能不再适用,需考虑数据分布的动态特性,有符号对称量化可作为兼顾性能与速度的优选方案。
  • 硬件加速兼容性:该方案保持了零点为零的特性,意味着可以直接复用现有的对称量化硬件加速内核,无需修改底层算子实现,极大地降低了工程落地门槛。
  • 模型部署趋势:随着端侧部署对延迟和内存的严苛要求,能够“零成本”提升精度的量化算法将成为模型服务化(Serving)阶段的重要竞争力指标。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Quantization 量化 Research 科学研究 Inference 推理