forceinjection.github.io/02_npu_programming/14_quantization/quantization_guide.html at main · ForceInjection/forceinjection.github.io · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
<!DOCTYPE html>
<html lang="zh">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>量化推理 — 交互式学习指南</title>
<style>
:root{--bg:#0d1117;--card:#161b22;--bdr:#30363d;--tx:#c9d1d9;--ac:#58a6ff;--gn:#3fb950;--or:#d2991d;--rd:#f85149;--dim:#8b949e}
*{margin:0;padding:0;box-sizing:border-box}
body{font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',sans-serif;background:var(--bg);color:var(--tx);padding:24px;max-width:1100px;margin:0 auto}
h1{font-size:26px;margin-bottom:4px}
.sub{color:var(--dim);font-size:14px;margin-bottom:28px}
h2{font-size:20px;margin:32px 0 16px;padding-bottom:8px;border-bottom:1px solid var(--bdr);color:var(--ac)}
h3{font-size:16px;margin:20px 0 10px}
.card{background:var(--card);border:1px solid var(--bdr);border-radius:8px;padding:20px;margin-bottom:20px}
table{width:100%;border-collapse:collapse;margin:12px 0;font-size:14px}
th,td{padding:8px 12px;text-align:left;border-bottom:1px solid var(--bdr)}
th{color:var(--ac);font-weight:600;font-size:13px}
td{font-size:14px}
.calc-row{font-family:'SF Mono',monospace;font-size:13px}
.step{color:var(--ac);font-weight:600}
.val-good{color:var(--gn)}
.val-bad{color:var(--rd)}
.val-dim{color:var(--dim)}
.val-hi{color:var(--or)}
.highlight{background:rgba(88,166,255,0.08);border-radius:4px;padding:2px 6px}
.code{background:#0d1117;border:1px solid var(--bdr);border-radius:6px;padding:16px;margin:12px 0;font-family:'SF Mono',monospace;font-size:13px;line-height:1.6;overflow-x:auto;white-space:pre}
.note{background:rgba(88,166,255,0.08);border-left:3px solid var(--ac);padding:10px 14px;margin:12px 0;border-radius:0 6px 6px 0;font-size:14px}
.warn{background:rgba(210,153,29,0.08);border-left:3px solid var(--or)}
.good{background:rgba(63,185,80,0.08);border-left:3px solid var(--gn)}
.bar-container{display:flex;gap:2px;align-items:flex-end;height:80px;margin:12px 0}
.bar{border-radius:3px 3px 0 0;display:flex;align-items:flex-end;justify-content:center;font-size:11px;padding-bottom:4px;color:#fff;font-weight:600}
.grid-2{display:grid;grid-template-columns:1fr 1fr;gap:16px}
.tag{display:inline-block;font-size:11px;padding:2px 8px;border-radius:10px;margin-right:4px}
.tag-g{background:rgba(63,185,80,0.15);color:var(--gn)}
.tag-r{background:rgba(248,81,73,0.15);color:var(--rd)}
.tag-o{background:rgba(210,153,29,0.15);color:var(--or)}
.calc-steps{font-size:13px;line-height:1.8}
.comparison td:nth-child(2){border-left:2px solid var(--ac)}
.comparison td:nth-child(4){border-left:2px solid var(--gn)}
nav{position:sticky;top:0;background:var(--bg);padding:12px 0;border-bottom:1px solid var(--bdr);margin-bottom:24px;display:flex;gap:16px;flex-wrap:wrap;z-index:10}
nav a{color:var(--dim);text-decoration:none;font-size:13px}
nav a:hover{color:var(--ac)}
</style>
</head>
<body>

<nav>
  <a href="#symmetric">§2 对称量化</a>
  <a href="#asymmetric">§3 非对称量化</a>
  <a href="#perchannel">§4 Per-Channel</a>
  <a href="#int4">§6 INT4</a>
  <a href="#memory">§7 加速原理</a>
</nav>

<h1>量化推理 — 交互式学习指南</h1>
<p class="sub">基于手算示例，逐步理解对称/非对称/Per-Channel/INT4 量化原理</p>

<!-- §2 对称量化 -->
<h2 id="symmetric">§2 对称量化 (Symmetric INT8)</h2>

<div class="card">
  <h3>2.1 数学公式</h3>
  <div class="code">量化:   q = round(x / scale),  clamp(q, -127, 127)
反量化: x' = q × scale
其中:   scale = max(|x|) / 127</div>
  <p style="font-size:14px;color:var(--dim);margin-top:8px">零点固定为 0——0.0 在浮点和量化空间都对应 0</p>
</div>

<div class="card">
  <h3>2.2 手算示例：权重 <code>[0.03, -0.05, 0.12, -0.08, 0.01]</code></h3>

  <div class="calc-steps">
    <p><span class="step">Step 1:</span> max(|x|) = <span class="val-hi">0.12</span></p>
    <p><span class="step">Step 2:</span> scale = 0.12 / 127 = <span class="val-hi">0.000945</span></p>
    <p style="margin-top:8px"><span class="step">Step 3: 量化 (÷ scale → round)</span></p>
  </div>

  <table style="margin-top:8px">
    <tr><th>原值</th><th>计算</th><th>INT8</th><th>反量化</th><th>误差</th></tr>
    <tr class="calc-row"><td> 0.03</td><td>round( 31.7)</td><td class="val-good"> 32</td><td> 0.0302</td><td class="val-dim">+0.0002</td></tr>
    <tr class="calc-row"><td>-0.05</td><td>round(-52.9)</td><td class="val-good">-53</td><td>-0.0501</td><td class="val-dim">-0.0001</td></tr>
    <tr class="calc-row"><td> 0.12</td><td>round(127.0)</td><td class="val-hi">127</td><td> 0.1200</td><td class="val-good">0</td></tr>
    <tr class="calc-row"><td>-0.08</td><td>round(-84.7)</td><td class="val-good">-85</td><td>-0.0803</td><td class="val-dim">-0.0003</td></tr>
    <tr class="calc-row"><td> 0.01</td><td>round( 10.6)</td><td class="val-good"> 11</td><td> 0.0104</td><td class="val-dim">+0.0004</td></tr>
  </table>
</div>

<div class="card">
  <h3>2.3 Outlier 的影响</h3>
  <p style="font-size:14px;margin-bottom:12px">999 个值在 [-0.5, 0.5]，<span class="val-hi">1 个 outlier = 5.0</span></p>

  <div class="grid-2">
    <div>
      <h4 style="font-size:14px;color:var(--gn)">无 outlier</h4>
      <div class="calc-steps">
        scale = 0.5/127 = <span class="val-good">0.00394</span><br>
        值 0.01 → q ≈ <span class="val-good">3</span><br>
        值 0.05 → q ≈ <span class="val-good">13</span><br>
        值 0.10 → q ≈ <span class="val-good">25</span>
      </div>
    </div>
    <div>
      <h4 style="font-size:14px;color:var(--rd)">有 outlier (5.0)</h4>
      <div class="calc-steps">
        scale = 5.0/127 = <span class="val-bad">0.0394</span> (×10)<br>
        值 0.01 → q = <span class="val-bad">0</span> ← 完全丢失<br>
        值 0.05 → q = <span class="val-bad">1</span><br>
        值 0.10 → q = <span class="val-bad">3</span>
      </div>
    </div>
  </div>

  <div class="note warn" style="margin-top:12px">
    <strong>1 个 outlier 放大 scale 10 倍</strong>，256 级精度只用了不到 4 级。<br>
    解决方案：LLM.int8() 对 outlier 维度单独保留 FP16，其余 INT8。
  </div>
</div>

<!-- §3 非对称量化 -->
<h2 id="asymmetric">§3 非对称量化 (Asymmetric UINT8)</h2>

<div class="card">
  <h3>3.1 数学公式</h3>
  <div class="code">量化:   q = round(x / scale) + zero_point,  clamp(q, 0, 255)
反量化: x' = (q - zero_point) × scale
其中:   scale = (x_max - x_min) / 255
        zero_point = round(-x_min / scale)</div>
  <p style="font-size:14px;color:var(--dim);margin-top:8px">zero_point 将数据范围精准映射到 [0, 255]，无浪费</p>
</div>

<div class="card">
  <h3>3.2 手算示例：ReLU 激活值 <code>[0.0, 0.35, 1.20, 0.08, 2.50]</code></h3>

  <div class="calc-steps">
    <p><span class="step">Step 1:</span> x_min = <span class="val-dim">0.0</span>, x_max = <span class="val-hi">2.50</span></p>
    <p><span class="step">Step 2:</span> scale = 2.50/255 = <span class="val-hi">0.00980</span>, zero_point = <span class="val-dim">0</span></p>
  </div>

  <table style="margin-top:8px">
    <tr><th>原值</th><th>计算</th><th>UINT8</th><th>反量化</th><th>误差</th></tr>
    <tr class="calc-row"><td>0.00</td><td>round(0.0)+0</td><td class="val-dim">  0</td><td>0.0000</td><td class="val-good">0</td></tr>
    <tr class="calc-row"><td>0.35</td><td>round(35.7)+0</td><td class="val-good"> 36</td><td>0.3528</td><td class="val-dim">+0.0028</td></tr>
    <tr class="calc-row"><td>1.20</td><td>round(122.4)+0</td><td class="val-good">122</td><td>1.1960</td><td class="val-dim">-0.004</td></tr>
    <tr class="calc-row"><td>0.08</td><td>round(8.2)+0</td><td class="val-good">  8</td><td>0.0784</td><td class="val-dim">-0.0016</td></tr>
    <tr class="calc-row"><td>2.50</td><td>round(255.1)+0</td><td class="val-hi">255</td><td>2.5000</td><td class="val-good">0</td></tr>
  </table>
</div>

<div class="card">
  <h3>3.3 同一组数据：对称 vs 非对称</h3>

  <table class="comparison">
    <tr><th style="width:16%">原值</th><th colspan="2" style="text-align:center">对称量化 (不适合)</th><th colspan="2" style="text-align:center">非对称量化 (适合)</th></tr>
    <tr><th></th><th style="color:var(--rd)">q → 反量化</th><th>误差</th><th style="color:var(--gn)">q → 反量化</th><th>误差</th></tr>
    <tr class="calc-row"><td>0.00</td><td>  0→0.0000</td><td class="val-good">0</td><td>  0→0.0000</td><td class="val-good">0</td></tr>
    <tr class="calc-row"><td>0.35</td><td class="val-bad"> 18→0.3544</td><td class="val-bad">Δ0.0044</td><td class="val-good"> 36→0.3528</td><td class="val-good">Δ0.0028</td></tr>
    <tr class="calc-row"><td>1.20</td><td> 61→1.2011</td><td class="val-dim">Δ0.0011</td><td>122→1.1960</td><td class="val-dim">Δ0.0040</td></tr>
    <tr class="calc-row"><td>0.08</td><td class="val-bad">  4→0.0788</td><td>Δ0.0012</td><td class="val-good">  8→0.0784</td><td>Δ0.0016</td></tr>
    <tr class="calc-row"><td>2.50</td><td>127→2.5000</td><td class="val-good">0</td><td>255→2.5000</td><td class="val-good">0</td></tr>
  </table>

  <div class="note good" style="margin-top:12px">
    <strong>对称浪费 50% 范围</strong>（负数区间 [-2.5,0) 用不上），非对称 <strong>scale 减半 → 精度翻倍</strong>
  </div>
</div>

<!-- §4 Per-Channel -->
<h2 id="perchannel">§4 Per-Tensor vs Per-Channel</h2>

<div class="card">
  <h3>4.1 粒度对比</h3>

  <table>
    <tr><th>方式</th><th>粒度</th><th>scale 数 (4096²)</th><th>存储开销</th><th>精度</th></tr>
    <tr><td><span class="tag tag-r">Per-Tensor</span></td><td>整个矩阵 1 个 scale</td><td>1</td><td>4 B</td><td>受 outlier 通道影响</td></tr>
    <tr><td><span class="tag tag-g">Per-Channel</span></td><td>每行独立 scale</td><td>4096</td><td>16 KB</td><td>各通道自主最优</td></tr>
  </table>
</div>

<div class="card">
  <h3>4.2 手算示例：2×3 矩阵</h3>
  <div class="code">通道 0 (小值): [ 0.05, -0.02,  0.10]
通道 1 (大值): [ 1.20, -0.80, -2.50]</div>

  <div class="grid-2">
    <div>
      <h4 style="color:var(--rd)">Per-Tensor</h4>
      <div class="calc-steps">
        scale = 2.50/127 = <span class="val-bad">0.01969</span><br><br>
        <strong>通道 0:</strong><br>
        0.05→3 → 0.059 <span class="val-bad">Δ+0.009 (18%!)</span><br>
        -0.02→-1 → -0.020 Δ0<br>
        0.10→5 → 0.098 Δ-0.002<br>
        RMSE = <span class="val-bad">0.0054</span><br><br>
        <strong>通道 1:</strong><br>
        1.20→61 → 1.201 Δ+0.001<br>
        -0.80→-41 → -0.807 Δ-0.007<br>
        -2.50→-127 → -2.501 Δ-0.001<br>
        RMSE = <span class="val-dim">0.0038</span>
      </div>
    </div>
    <div>
      <h4 style="color:var(--gn)">Per-Channel</h4>
      <div class="calc-steps">
        <strong>通道 0:</strong> scale₀ = 0.10/127 = <span class="val-good">0.000787</span><br>
        0.05→64 → 0.0504 <span class="val-good">Δ+0.0004</span><br>
        -0.02→-25 → -0.0197 Δ+0.0003<br>
        0.10→127 → 0.1000 <span class="val-good">Δ0</span><br>
        RMSE = <span class="val-good">0.0003</span><br><br>
        <strong>通道 1:</strong> scale₁ = 2.50/127 = 0.01969<br>
        (与 per-tensor 相同)<br>
        RMSE = <span class="val-dim">0.0038</span><br><br>
        <span class="val-good">通道 0 精度提升 18×</span><br>
        额外开销仅 <span class="val-dim">8 bytes</span>
      </div>
    </div>
  </div>
</div>

<!-- §6 INT4 -->
<h2 id="int4">§6 INT4 量化</h2>

<div class="card">
  <h3>6.1 INT4 vs INT8</h3>
  <table>
    <tr><th></th><th>离散值</th><th>精度</th><th>7B HBM</th></tr>
    <tr><td>INT8</td><td>256</td><td>≈ 0.4%</td><td>~7 GB</td></tr>
    <tr><td>INT4</td><td>16</td><td>≈ 6.25%</td><td>~3.5 GB</td></tr>
  </table>
</div>

<div class="card">
  <h3>6.2 手算示例：INT4 Per-Tensor vs Group-Wise</h3>
  <p style="font-size:14px;margin-bottom:12px">8 个权重，INT4 范围 [-7, 7]，16 个离散级别</p>

  <div class="code">前 4 个 (小值): [0.15, -0.08, 0.03, 0.12]
后 4 个 (大值): [2.10, -1.80, 1.50, -0.95]</div>

  <div class="grid-2" style="margin-top:12px">
    <div>
      <h4 style="color:var(--rd)">Per-Tensor INT4</h4>
      <div class="calc-steps">
        scale = 2.10/7 = <span class="val-bad">0.30</span><br><br>
        <strong>前 4 个:</strong><br>
        0.15→1 → 0.30 <span class="val-bad">(原值 0.15，偏差极大)</span><br>
        -0.08→0 → <span class="val-bad">0</span><br>
        0.03→0 → <span class="val-bad">0</span><br>
        0.12→0 → <span class="val-bad">0</span><br>
        <span class="val-bad">全部塌缩到 0-1!</span><br>
        RMSE = <span class="val-bad">0.126</span><br><br>
        <strong>后 4 个:</strong><br>
        2.10→7 ✓<br>RMSE = <span class="val-dim">0.039</span>
      </div>
    </div>
    <div>
      <h4 style="color:var(--gn)">Group-Wise INT4</h4>
      <div class="calc-steps">
        <strong>Group 0:</strong> scale₀ = 0.15/7 = <span class="val-good">0.0214</span><br>
        0.15→7 → 0.150 <span class="val-good">✓</span><br>
        -0.08→-4 → -0.086 ✓<br>
        0.03→1 → 0.021 ✓<br>
        0.12→6 → 0.128 ✓<br>
        RMSE = <span class="val-good">0.011</span><br><br>
        <strong>Group 1:</strong> scale₁ = 2.10/7 = 0.30<br>
        (与 per-tensor 相同)<br>
        RMSE = <span class="val-dim">0.039</span><br><br>
        <span class="val-good">精度恢复！</span><br>
        开销: 2 scale × 4B = <span class="val-dim">8B</span>
      </div>
    </div>
  </div>

  <div class="note good" style="margin-top:12px">
    <strong>Group-wise 用 ~3% 额外存储</strong>，换回 per-channel 级别的精度。INT4 必须搭配 group-wise。
  </div>
</div>

<!-- §7 Memory -->
<h2 id="memory">§7 量化加速原理</h2>

<div class="card">
  <h3>7.1 显存节省 (7B 模型)</h3>
  <div class="bar-container">
    <div class="bar" style="width:120px;height:100%;background:var(--ac)">FP32<br>28GB</div>
    <div class="bar" style="width:100px;height:50%;background:var(--gn)">FP16<br>14GB</div>
    <div class="bar" style="width:60px;height:25%;background:var(--or)">INT8<br>7GB</div>
    <div class="bar" style="width:40px;height:12.5%;background:var(--rd)">INT4<br>3.5GB</div>
  </div>
</div>

<div class="card">
  <h3>7.2 HBM 带宽节省</h3>
  <p style="font-size:14px;margin-bottom:12px">推理时需从 HBM 读取全部权重。910B3 HBM 带宽 1538 GB/s。</p>
  <table>
    <tr><th>精度</th><th>权重大小</th><th>读取时间</th><th>加速</th></tr>
    <tr class="calc-row"><td>FP16</td><td>14 GB</td><td>9.1 ms</td><td>1×</td></tr>
    <tr class="calc-row"><td>INT8</td><td>7 GB</td><td>4.6 ms</td><td class="val-good">2×</td></tr>
    <tr class="calc-row"><td>INT4</td><td>3.5 GB</td><td>2.3 ms</td><td class="val-good">4×</td></tr>
  </table>
</div>

<div class="card" style="text-align:center;color:var(--dim);font-size:13px">
  更多交互式实验请打开 <code>quantization_viz.html</code> | Python 演示：<code>python3 quantization_demo.py</code>
</div>

</body>
</html>