ROSE-Webpage/index.html at master · mingluo-su/ROSE-Webpage · 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
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
<!DOCTYPE html>
<html lang="en">

<head>
  <meta charset="utf-8">
  <meta name="viewport" content="width=device-width, initial-scale=1">

  <!-- Primary Meta Tags -->
  <meta name="title" content="ROSE: Reordered SparseGPT for More Accurate One-Shot Large Language Models Pruning">
  <meta name="description"
    content="ROSE is a reordered one-shot pruning method that improves SparseGPT by adaptively reordering weight columns to minimize reconstruction error in large language models.">
  <meta name="keywords" content="Large Language Models, Pruning, One-Shot, Training-Free, SparseGPT, Model Compression">
  <meta name="author" content="Mingluo Su, Huan Wang">
  <meta name="robots" content="index, follow">
  <meta name="language" content="English">

  <!-- Open Graph / Facebook -->
  <meta property="og:type" content="article">
  <meta property="og:site_name" content="Westlake University">
  <meta property="og:title"
    content="ROSE: Reordered SparseGPT for More Accurate One-Shot Large Language Models Pruning">
  <meta property="og:description"
    content="ROSE is a reordered one-shot pruning method that improves SparseGPT by adaptively reordering weight columns to minimize reconstruction error in large language models.">
  <meta property="og:url" content="https://mingluo-su.github.io/ROSE/">
  <meta property="og:image" content="https://mingluo-su.github.io/ROSE/static/images/social_preview.png">
  <meta property="og:image:width" content="1200">
  <meta property="og:image:height" content="630">
  <meta property="og:image:alt" content="ROSE - Research Overview">
  <meta property="article:published_time" content="2025-01-01T00:00:00.000Z">
  <meta property="article:author" content="Mingluo Su, Huan Wang">
  <meta property="article:section" content="Research">
  <meta property="article:tag" content="Pruning">
  <meta property="article:tag" content="Large Language Models">

  <!-- Twitter -->
  <meta name="twitter:card" content="summary_large_image">
  <meta name="twitter:site" content="@westlakeuni">
  <meta name="twitter:creator" content="@huanwang_tech">
  <meta name="twitter:title"
    content="ROSE: Reordered SparseGPT for More Accurate One-Shot Large Language Models Pruning">
  <meta name="twitter:description"
    content="ROSE is a reordered one-shot pruning method that improves SparseGPT by adaptively reordering weight columns to minimize reconstruction error in large language models.">
  <meta name="twitter:image" content="https://mingluo-su.github.io/ROSE/static/images/social_preview.png">
  <meta name="twitter:image:alt" content="ROSE - Research Overview">

  <!-- Academic/Research Specific -->
  <meta name="citation_title"
    content="ROSE: Reordered SparseGPT for More Accurate One-Shot Large Language Models Pruning">
  <meta name="citation_author" content="Su, Mingluo">
  <meta name="citation_author" content="Wang, Huan">
  <meta name="citation_publication_date" content="2025">
  <meta name="citation_conference_title" content="Preprint">
  <meta name="citation_pdf_url" content="https://arxiv.org/pdf/2510.06751">

  <!-- Additional SEO -->
  <meta name="theme-color" content="#2563eb">
  <meta name="msapplication-TileColor" content="#2563eb">
  <meta name="apple-mobile-web-app-capable" content="yes">
  <meta name="apple-mobile-web-app-status-bar-style" content="default">

  <!-- Preconnect for performance -->
  <link rel="preconnect" href="https://fonts.googleapis.com">
  <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
  <link rel="preconnect" href="https://ajax.googleapis.com">
  <link rel="preconnect" href="https://documentcloud.adobe.com">
  <link rel="preconnect" href="https://cdn.jsdelivr.net">

  <title>ROSE: Reordered SparseGPT for More Accurate One-Shot Large Language Models Pruning</title>

  <!-- Favicon and App Icons -->
  <link rel="icon" type="image/x-icon" href="static/images/favicon.ico">
  <link rel="apple-touch-icon" href="static/images/favicon.ico">

  <!-- Critical CSS - Load synchronously -->
  <link rel="stylesheet" href="static/css/bulma.min.css">
  <link rel="stylesheet" href="static/css/index.css">

  <!-- Non-critical CSS - Load asynchronously -->
  <link rel="preload" href="static/css/bulma-carousel.min.css" as="style"
    onload="this.onload=null;this.rel='stylesheet'">
  <link rel="preload" href="static/css/bulma-slider.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
  <link rel="preload" href="static/css/fontawesome.all.min.css" as="style"
    onload="this.onload=null;this.rel='stylesheet'">
  <link rel="preload" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css" as="style"
    onload="this.onload=null;this.rel='stylesheet'">

  <!-- Fallback for browsers that don't support preload -->
  <noscript>
    <link rel="stylesheet" href="static/css/bulma-carousel.min.css">
    <link rel="stylesheet" href="static/css/bulma-slider.min.css">
    <link rel="stylesheet" href="static/css/fontawesome.all.min.css">
    <link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
  </noscript>

  <!-- Fonts -->
  <link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&display=swap" rel="stylesheet">

  <!-- Defer non-critical JavaScript -->
  <script defer src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
  <script defer src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
  <script defer src="static/js/fontawesome.all.min.js"></script>
  <script defer src="static/js/bulma-carousel.min.js"></script>
  <script defer src="static/js/bulma-slider.min.js"></script>
  <script defer src="static/js/index.js"></script>

  <!-- Structured Data for Academic Papers -->
  <script type="application/ld+json">
  {
    "@context": "https://schema.org",
    "@type": "ScholarlyArticle",
    "headline": "ROSE: Reordered SparseGPT for More Accurate One-Shot Large Language Models Pruning",
    "description": "ROSE is a reordered one-shot pruning method that improves SparseGPT by adaptively reordering weight columns to minimize reconstruction error in large language models.",
    "author": [
      {
        "@type": "Person",
        "name": "Mingluo Su",
        "affiliation": {
          "@type": "Organization",
          "name": "Westlake University"
        }
      },
      {
        "@type": "Person",
        "name": "Huan Wang",
        "affiliation": {
          "@type": "Organization",
          "name": "Westlake University"
        }
      }
    ],
    "datePublished": "2025-01-01",
    "publisher": {
      "@type": "Organization",
      "name": "arXiv"
    },
    "url": "https://mingluo-su.github.io/ROSE/",
    "image": "https://mingluo-su.github.io/ROSE/static/images/social_preview.png",
    "keywords": ["Large Language Models", "Pruning", "One-Shot", "SparseGPT", "Model Compression"],
    "abstract": "Pruning is widely recognized as an effective method for reducing the parameters of large language models (LLMs), potentially leading to more efficient deployment and inference. One classic and prominent path of LLM one-shot pruning is to leverage second-order gradients (i.e., Hessian), represented by the pioneering work SparseGPT. However, the predefined left-to-right pruning order in SparseGPT leads to suboptimal performance when the weights exhibit columnar patterns. This paper studies the effect of pruning order under the SparseGPT framework. The analyses lead us to propose ROSE, a reordered SparseGPT method that prioritizes weights with larger potential pruning errors to be pruned earlier. ROSE first performs pre-pruning to identify candidate weights for removal, and estimates both column and block pruning loss. Subsequently, two-level reordering is performed: columns within each block are reordered in descending order of column loss, while blocks are reordered based on block loss. We introduce the relative range of block loss as a metric to identify columnar layers, enabling adaptive reordering across the entire model. Substantial empirical results on prevalent LLMs (LLaMA2-7B/13B/70B, LLaMA3-8B, Mistral-7B) demonstrate that ROSE surpasses the original SparseGPT and other counterpart pruning methods..",
    "isAccessibleForFree": true,
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "mainEntity": {
      "@type": "WebPage",
      "@id": "https://mingluo-su.github.io/ROSE/"
    },
    "about": [
      { "@type": "Thing", "name": "Model Compression" },
      { "@type": "Thing", "name": "Large Language Models" }
    ]
  }
  </script>

  <!-- Website/Organization Structured Data -->
  <script type="application/ld+json">
  {
    "@context": "https://schema.org",
    "@type": "Organization",
    "name": "Westlake University",
    "url": "https://mingluo-su.github.io/ROSE/",
    "logo": "https://mingluo-su.github.io/ROSE/static/images/favicon.ico",
    "sameAs": [
      "https://twitter.com/huanwang_tech",
      "https://github.com/mingluo-su"
    ]
  }
  </script>

  <style>
    .publication-authors {
      margin-bottom: 0.2rem;
    }

    .bibtex-header {
      display: flex;
      justify-content: space-between;
      align-items: center;
    }

    .copy-bibtex-btn {
      background: none;
      border: none;
      cursor: pointer;
      font-size: 1rem;
      color: #2563eb;
    }
  </style>
</head>

<body>

  <!-- Scroll to Top Button -->
  <button class="scroll-to-top" onclick="scrollToTop()" title="Scroll to top" aria-label="Scroll to top">
    <i class="fas fa-chevron-up"></i>
  </button>

  <main id="main-content">
    <section class="hero">
      <div class="hero-body" style="padding-bottom: 1rem;">
        <div class="container is-max-desktop">
          <div class="columns is-centered">
            <div class="column has-text-centered">
              <h1 class="title is-1 publication-title" style="margin-bottom: 0.5rem;">
                ROSE: Reordered SparseGPT for More Accurate One-Shot Large Language Models Pruning
              </h1>

              <div class="is-size-3 has-text-weight-bold" style="color: #b31b1b; margin-bottom: 1.5rem;">
                CPAL 2026 oral
              </div>

              <div class="is-size-5 publication-authors">
                <span class="author-block">
                  <a href="https://mingluo-su.github.io" target="_blank">Mingluo Su</a><sup>1</sup>,
                </span>
                <span class="author-block">
                  <a href="https://huanwang.tech/" target="_blank">Huan Wang</a><sup>1*</sup>
                </span>
              </div>

              <div class="is-size-5 publication-authors">
                <span class="author-block"><sup>1</sup>Westlake University</span>
              </div>

              <div class="is-size-7" style="margin-top: 0.5rem; margin-bottom: 1.5rem;">
                <span class="author-block"><sup>*</sup>Corresponding author: wanghuan [at] westlake [dot] edu [dot]
                  cn</span>
              </div>

              <div class="columns is-centered logo-stack mb-0 pb-0">
                <div class="logo-row">
                  <div class="column is-narrow">
                    <img src="static/images/westlake.png" alt="Westlake University" style="height: 100px;">
                  </div>
                  <div class="column is-narrow">
                    <img src="static/images/logo.jpg" alt="ENCODE LAB" style="height: 100px;">
                  </div>
                </div>
              </div>

              <div class="column has-text-centered">
                <div class="publication-links">
                  <span class="link-block">
                    <a href="https://arxiv.org/abs/2603.05878" target="_blank"
                      class="external-link button is-normal is-rounded is-dark">
                      <span class="icon"><i class="fas fa-file-pdf"></i></span>
                      <span>Paper</span>
                    </a>
                  </span>
                  <span class="link-block">
                    <a href="https://github.com/mingluo-su/ROSE" target="_blank"
                      class="external-link button is-normal is-rounded is-dark">
                      <span class="icon"><i class="fab fa-github"></i></span>
                      <span>Code</span>
                    </a>
                  </span>
                </div>
              </div>
            </div>
          </div>
        </div>
      </div>
    </section>
  </main>

  <section class="section hero is-light">
    <div class="container is-max-desktop">
      <div class="columns is-centered has-text-centered">
        <div class="column is-four-fifths">
          <h2 class="title is-3">Abstract</h2>
          <div class="content has-text-justified is-size-6">
            <p>
              Pruning is widely recognized as an effective method for reducing the parameters of large language models (LLMs), potentially leading to more efficient deployment and inference. One classic and prominent path of one-shot LLM pruning is to leverage second-order gradients (<i>i.e.</i>, Hessian), represented by the pioneering work SparseGPT. However, the predefined left-to-right pruning order in SparseGPT leads to suboptimal performance when the weights exhibit <i>columnar</i> patterns. This paper studies the effect of pruning order under the SparseGPT framework. The analyses lead us to propose ROSE, a reordered SparseGPT method that prioritizes weight columns with larger potential pruning errors to be processed earlier. ROSE first performs pre-pruning to identify weights that are highly likely to be pruned, and estimates both column and block pruning loss. Subsequently, two-level reordering is performed: columns within each block are reordered in descending order of column loss, while blocks are reordered based on block loss. We introduce the relative range of block loss as a metric to identify columnar layers, enabling adaptive reordering across the entire model. Substantial empirical results on prevalent LLMs (LLaMA2-7B/13B/70B,
              LLaMA3-8B, Mistral-7B) demonstrate that ROSE surpasses the original SparseGPT and other counterpart
              pruning methods.
            </p>
          </div>
        </div>
      </div>
    </div>
  </section>

  <section class="section hero is-light">
    <div class="container is-max-desktop">
      <div class="columns is-centered has-text-centered">
        <div class="column is-four-fifths">
          <h2 class="title is-3">Motivation</h2>
          <div class="content">
            <img src="static/images/motivation.png" alt="Motivation">
          </div>
          <h2 class="subtitle is-size-6" style="text-align: left;">
            (a) Change in reconstruction error of the "self\_attn.o\_proj" layer in the first Transformer Block of
            LLaMA2-7B during SparseGPT pruning as the number of pruned blocks increases. The sharpest increase appears
            at a later stage. (b) Weight visualization shows a <i>columnar</i> pattern along the input channel, with one
            block containing the most concentrated high-magnitude weights. (c) Reconstruction error after reordering:
            pruning the high-error block earlier yields lower total error.
          </h2>
        </div>
      </div>
    </div>
  </section>


  <section class="section hero is-light">
    <div class="container is-max-desktop">
      <div class="columns is-centered has-text-centered">
        <div class="column is-four-fifths">
          <h2 class="title is-3">Overview of our <i>ROSE</i> method</h2>
          <div class="content">
            <img src="static/images/overview.png" alt="Overview of the ROSE">
          </div>
          <h2 class="subtitle is-size-6" style="text-align: left;">
            (a) Overview of the difference between SparseGPT and ROSE. In SparseGPT, as pruning progresses, fewer
            weights remain for compensation. If high-error weights are pruned late, compensation is limited. ROSE
            reorders columns to make high-error weights be pruned early, preserving more parameters for compensation. (b) Illustration of our ROSE for the <i>columnar</i> layer. Given the dense weight <b> W</b> and target sparsity rate <i>p</i>%, we calculate the importance score <b>S</b> and split it into blocks based on blocksize.The smallest <i>p</i>% of values from each block are selected as the loss matrix <b>L</b>.  Column loss and block loss are calculated based on the loss matrix. Columns within one block are reordered in descending order of column loss, and blocks are reordered in descending order of block loss.
.
          </h2>
        </div>
      </div>
    </div>
  </section>

  <!-- Image carousel -->
  <section class="hero is-small" style="margin-bottom: 4rem;">
    <div class="hero-body">
      <div class="container is-max-desktop">

        <h2 class="title is-3 has-text-centered" style="margin-bottom: 0rem;">Main Results</h2>
        <div id="results-carousel" class="carousel results-carousel" style="margin-top: -3rem;">
          <div class="item"
            style="min-height: 100%; display: flex; flex-direction: column; justify-content: center; align-items: center;">
            <img src="static/images/error.png" alt="Relative reconstruction error" loading="lazy" />
            <h2 class="subtitle has-text-centered">
              Relative reconstruction error of the "self\_attn.o\_proj" layer in the second Transformer Block of
              LLaMA2-7B by ROSE and its variants at varying sparsity rates.
            </h2>
          </div>

          <div class="item"
            style="min-height: 100%; display: flex; flex-direction: column; justify-content: center; align-items: center;">
            <img src="static/images/result-ppl.png" alt="Second research result visualization" loading="lazy" />
            <h2 class="subtitle has-text-centered">
              WikiText perplexity performance on LLaMA3-8B and Mistral-7B at varying sparsity rates. The best results
              are highlighted in <b>bold</b> and the second-best results are indicated with <u>underline</u>.
            </h2>
          </div>

          <div class="item">
            <img src="static/images/result-model.png" alt="Third research result visualization"
              style="max-width: 100%; height: auto;" loading="lazy" />
            <h2 class="subtitle has-text-centered">
              WikiText perplexity and zero-shot accuracy on LLaMA2 models at 70% sparsity for different unstructured
              pruning methods. The best results are highlighted in <b>bold</b> and the second-best results are indicated
              with <u>underline</u>.
            </h2>
          </div>


        </div>
      </div>
    </div>
  </section>
  <!-- End image carousel -->

  <section class="section" id="BibTeX">
    <div class="container is-max-desktop content">
      <div class="bibtex-header">
        <h2 class="title">BibTeX</h2>
        <button class="copy-bibtex-btn" onclick="copyBibTeX()" title="Copy BibTeX to clipboard">
          <i class="fas fa-copy"></i>
          <span class="copy-text">Copy</span>
        </button>
      </div>
      <pre id="bibtex-code"><code>@inproceedings{su2025rose,
  title={ROSE: Reordered SparseGPT for More Accurate One-Shot Large Language Models Pruning},
  author={Su, Mingluo and Wang, Huan},
  booktitle={CPAL},
  year={2026}
}</code></pre>
    </div>
  </section>

  <footer class="footer">
    <div class="container">
      <div class="columns is-centered">
        <div class="column is-8">
          <div class="content has-text-centered">
            <a href="https://mapmyvisitors.com/web/1c2cf" title="Visit tracker">
              <img src="https://mapmyvisitors.com/map.png?d=U9tGOHPEQveyV-2QSJF-0ugyIQ7eUGEN1Fm50LLksvs&cl=ffffff"
                alt="Visitor map" />
            </a>
            <p>
              This page was built using the <a href="https://github.com/eliahuhorwitz/Academic-project-page-template"
                target="_blank">Academic Project Page Template</a>, adapted from the <a href="https://nerfies.github.io"
                target="_blank">Nerfies</a> project page.
            </p>
          </div>
        </div>
      </div>
    </div>
  </footer>

  <script>
    function copyBibTeX() {
      const code = document.getElementById('bibtex-code').innerText;
      navigator.clipboard.writeText(code).then(() => {
        const btn = document.querySelector('.copy-text');
        const original = btn.textContent;
        btn.textContent = 'Copied!';
        setTimeout(() => btn.textContent = original, 2000);
      });
    }
    function scrollToTop() {
      window.scrollTo({ top: 0, behavior: 'smooth' });
    }
  </script>

</body>

</html>