Auturnal gold A, leaf and orbit markAuturnalRESEARCH & PERSPECTIVES
← All research guides

01 / Model architecture

Transformer: parallel computation, long-range dependencies and the cost of attention

Original title:Transformer 的真正突破:并行计算、长程依赖与注意力的代价

From Transformer to FlashAttention: separate architecture, algorithmic complexity and hardware execution.

2026-10-16 · v4 · Compiled by Codex (AI) · About 5 min read

Transformer: parallel computation, long-range dependencies and the cost of attention: AI-generated conceptual cover art, not an experimental result.
Cover · Conceptual illustration generated with ImageGen; not an experimental result.

AI-assisted literature guide · Not independently peer reviewed. v3 adds language switching; claims and references retain the published original.

The body, charts and source material remain in their original Simplified Chinese. Language switching changes the interface, title and introduction only. Translation note: non-Chinese interface text, titles and introductions are AI-assisted translations without independent language review. The original and its sources remain available.

This review separates the architectural, algorithmic, and hardware contributions of attention models. Analytical memory examples illustrate scaling costs without claiming new model benchmarks.

摘要

Transformer 把序列位置之间的信息交换组织为注意力计算,为并行训练提供了不同于递归网络的路径。本文沿着“信息如何流动—计算如何增长—数据如何搬运”三条线索,分析原始注意力、稀疏注意力与 IO 优化的区别。核心判断是:架构允许并行、数学运算量和实际运行速度,应分别讨论,不能用一个吞吐数字替代。[1][2]

1. 先明确被改变的瓶颈

递归序列模型在一个样本内部,通常需要沿时间位置依次更新状态。原始 Transformer 用位置表示补充顺序信息,再让多个位置通过注意力交换内容。它缩短了远距离位置之间的信息路径,也增加了训练时跨位置并行的空间。不过,自回归解码仍需要依赖已经生成的词元;训练可以并行,不等于整段答案可以同时生成。[1]

2. Q、K、V 是计算角色

缩放点积注意力可以写成 :softmax⁡ ⁣(QKTdk)V\operatorname{softmax}\!\left(\frac{QK^{\mathsf T}}{\sqrt{d_k}}\right)VQ 与 K 形成匹配分数,softmax 将分数归一化,再对 V 加权。多头机制让不同投影子空间分别完成这一操作,再合并结果。这些名称描述向量在计算中的作用,并不意味着模型拥有可直接对应人类认知的“问题”“记忆”或“理解”模块。[1]

LaTeX source\operatorname{softmax}\!\left(\frac{QK^{\mathsf T}}{\sqrt{d_k}}\right)V

3. 性能数字必须保留实验语境

原论文英文到德文 WMT 2014 实验报告,Transformer big 达到 28.4 BLEU。这个数字属于特定数据、分词、训练、解码与评分设置,不是现代通用语言模型的能力证明。本文不比较不同论文的英法分数:可访问版本的摘要、表格与正文存在不同写法,混用会制造并不存在的精确结论。[1]

4. 长上下文为什么昂贵

完整注意力在长度为 n 的序列上建立 n×nn\times n 配对关系。图 2 只计算“单层、单头、显式保存 float32 分数矩阵”的字节数:n2×4n^2\times4;长度翻倍,这一部分变为四倍。它不是完整训练显存估计,未计入梯度、激活、投影、批量、多头与优化器,也不适用于从不完整落盘分数矩阵的实现。[1][2]

5. 两种优化回答不同的问题

Longformer 改变哪些位置需要交互,用局部窗口与选择性的全局注意力处理长文档。FlashAttention 则围绕存储层级重排计算,借助分块与融合减少高带宽显存访问,保持其所定义的精确注意力。前者调整连接结构;后者优化同一计算的执行方式。不能把 IO 优化叫成消除了所有二次计算量。[3][2]

图表:为显式注意力分数矩阵的解析存储量;仅单头、单层、float32,非完整显存或速度实测。
图 1 · 图 1 为显式注意力分数矩阵的解析存储量;仅单头、单层、float32,非完整显存或速度实测。

6. 可以复核的解析例子

取序列长度 512、1024、2048、4096,显式分数矩阵分别需要 1、4、16、64 MiB。计算使用二进制 MiB,1 MiB=2201\,\mathrm{MiB}=2^{20} 字节。图中的数字来自公式推算,不是 GPU 实测。若实现把分数分块留在片上存储,峰值占用会不同;因此应先看内核执行策略,再用这个例子解释趋势。[1][2]

7. 应如何比较一个“更快”的方案

本文建议分别报告训练样本长度、批量、精度、设备、预热、吞吐、首词元延迟和峰值显存,并在相同质量目标下比较。若一个方案截断了上下文,另一个保留全部输入,速度差包含任务变化;若一方使用融合内核,另一方使用朴素实现,差别也不能全部归因于架构。这是实验设计建议,并非本文的新实验结果。

一条容易忽略的计算边界

注意力分数的 n2n^2 项只描述位置之间的交互,投影与逐位置前馈仍有自己的计算成本。对于固定序列长度,增大隐藏维度会改变这些项的相对比重;对于固定维度,延长序列则会让位置配对更重要。因此,单凭“注意力是二次复杂度”不能预测整个模型每一部分的耗时占比。应先列出模型维度与输入,再解释实际剖析结果。

后续实验可以如何设计

可选择相同参数与序列,分别使用朴素完整注意力与已验证的融合实现,先检查输出与梯度在约定容差内一致,再测峰值显存、时间和吞吐。局部注意力应另设质量对照,因为它改变连接关系。预热、同步、精度与输入长度必须固定;首次编译和稳定运行应分别报告。该方案是待执行实验设计,本文没有生成对应测量记录。

局限与结论

这里讨论的是三篇基础研究揭示的机制,不能据此断言任一当前产品最快或最准确。注意力权重可以辅助观察信息交互,却不自动构成可信的因果解释。选择架构时,应把依赖范围、质量损失与设备约束同时列出;只有计算对象和测量口径一致,性能判断才有可比性。[1][3][2]

Reading and verification checklist
比较问题应保留的条件不能直接推出
架构并行性训练或自回归解码阶段整段输出同时生成
稀疏连接窗口、全局词元与质量精确保留全部交互
IO 优化内核、设备、精度与尺寸消除全部二次运算

Primary sources and claim mapping

  1. Vaswani et al. · Attention Is All You NeedOpen the primary source ↗Source date:2017-06-12;Checked:2026-10-15;Location:v7; sections 3-4; Table 2 (EN-DE only)Transformer uses scaled dot-product attention; full attention has quadratic pair complexity. EN-DE big model 28.4 BLEU; no cross-version EN-FR claim.
  2. Dao et al. · FlashAttention: Fast and Memory-Efficient Exact Attention with IO-AwarenessOpen the primary source ↗Source date:2022-05-27;Checked:2026-10-15;Location:v1; sections 2-3Exact attention reorganizes memory IO using tiling; this is not a linear-time approximation of attention.
  3. Beltagy et al. · Longformer: The Long-Document TransformerOpen the primary source ↗Source date:2020-04-10;Checked:2026-10-15;Location:v2 abstract and architecture descriptionLocal window attention combines with selected global attention; inference limits depend on architecture.
Figure provenance, data and assumptions
[
  {
    "slug": "transformer-attention",
    "role": "conceptual-cover",
    "generator": "OpenAI built-in image_gen",
    "ai_generated": true,
    "alt": "暖金细线连接深色玻璃平面的注意力主题概念插画。",
    "original_sha256": "9ae6e8d6e113d034f86d32e08dd248c561c3ee362928b63d8ea5df1ca2646a57",
    "prompt": "Create one exceptional premium science journal editorial COVER ART for an article about Transformer attention. Landscape 3:2 composition, edge-to-edge fine art, no border, NO TEXT, NO LETTERS, NO NUMBERS, NO LOGOS. A cinematic macro photograph of a real gallery-quality sculptural installation: many extremely thin brushed-gold threads span between three floating smoke-glass rectangular planes, some paths selectively illuminated, a beautiful woven fan of attention connections, precise engineering yet poetic. Make the sculpture fill the frame and read immediately at small thumbnail size. Deep ink-blue near-black background fading to slate, warm amber edge light, exquisite gold highlights, subtle ivory glass caustics, delicate dust, believable tactile material and physically grounded shadows, a striking asymmetrical diagonal composition with generous breathing room around the silhouette. Editorial art direction like a sophisticated contemporary scientific publication, museum photography, large-format lens, rich midtones, elegant restraint, immaculate craft. Not a chart, no scientific labels, not a literal experimental visualization. Avoid neon cyberpunk, generic glowing brain, humanoid robot, stock-tech chip imagery, flat infographic boxes, plastic CGI, rainbows, cheap gradients. This is an openly disclosed conceptual illustration of selective information connections.",
    "files": [
      {
        "file": "assets/covers/transformer-attention.png",
        "width": 1280,
        "height": 853,
        "bytes": 1559169,
        "sha256": "3d298d29aa96387052dcbe2558718f6711dd9b2d553ef98e778b9814686ad2be",
        "url": "/guides/assets/v2-transformer-attention-3d298d29aa96.png"
      },
      {
        "file": "assets/covers/transformer-attention-480.png",
        "width": 480,
        "height": 320,
        "bytes": 248011,
        "sha256": "b88782d3981eb0d0d6e2d00a76a0bc48c1fa4f66d23dc2b408b4b48619ef53bb",
        "url": "/guides/assets/v2-transformer-attention-480-b88782d3981e.png"
      }
    ],
    "url": "/guides/assets/v2-transformer-attention-3d298d29aa96.png",
    "thumbnail_url": "/guides/assets/v2-transformer-attention-480-b88782d3981e.png"
  },
  {
    "slug": "transformer-attention",
    "role": "data-chart",
    "file": "assets/charts/transformer-attention.png",
    "width": 1800,
    "height": 1125,
    "sha256": "13c38400f67a7c4c7adde30b5e03f64b3631e18c4975e3cb833b30254933ebd9",
    "bytes": 110862,
    "alt": "图表:为显式注意力分数矩阵的解析存储量;仅单头、单层、float32,非完整显存或速度实测。",
    "data": {
      "kind": "analytical",
      "length": [
        512,
        1024,
        2048,
        4096
      ],
      "MiB": [
        1.0,
        4.0,
        16.0,
        64.0
      ],
      "formula": "n*n*4/(2**20)",
      "excludes": "batch, heads, activations, gradients, optimizer; fused implementations differ"
    },
    "data_source_sha256": "963323d4a7a0cad4aebd986f09a55251a7c21f68c757d21b76245a128675e13e",
    "ai_generated": true,
    "disclosure": "AI-assisted deterministic rendering of unchanged v1 data; not a new experiment.",
    "url": "/guides/assets/v2-transformer-attention-13c38400f67a.png"
  }
]
Author, versions and corrections

v3 adds five interface languages while retaining the original text, 29 sources and chart data. v1 and v2 remain available.

文献导读与方法分析;不是原创实验论文、正式期刊录用或专业资质认证。解析示例不代表本站或当前产品的实测表现。

View original v1 · Download the article, figures and source mappingView original v2