-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
665 lines (627 loc) · 38.6 KB
/
index.html
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
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<!-- Meta tags for social media banners, these should be filled in appropriatly as they are your "business card" -->
<!-- Replace the content tag with appropriate information -->
<meta name="description" content="Efficient, Scalable and Generalizable LLM Fine-tuning with S²FT">
<meta property="og:title" content="S²FT" />
<meta property="og:description" content="Efficient, Scalable and Generalizable LLM Fine-tuning with S²FT" />
<meta property="og:url" content="https://infini-ai-lab.github.io/S2FT-Page/" />
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X630-->
<meta property="og:image" content="static/images/logo.png" />
<meta property="og:image:width" content="1200" />
<meta property="og:image:height" content="630" />
<meta name="twitter:title" content="S²FT">
<meta name="twitter:description" content="Efficient, Scalable and Generalizable LLM Fine-tuning with S²FT">
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X600-->
<meta name="twitter:image" content="static/images/logo.png">
<meta name="twitter:card" content="summary_large_image">
<!-- Keywords for your paper to be indexed by-->
<meta name="keywords" content="LLM; PEFT">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>S²FT: Efficient, Scalable and Generalizable LLM Fine-tuning by Structured Sparsity</title>
<link rel="icon" type="image/x-icon" href="static/images/logo.png">
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<link rel="stylesheet" href="static/css/bulma.min.css">
<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">
<link rel="stylesheet" href="static/css/index.css">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
<script defer src="static/js/fontawesome.all.min.js"></script>
<script src="static/js/bulma-carousel.min.js"></script>
<script src="static/js/bulma-slider.min.js"></script>
<script src="static/js/index.js"></script>
</head>
<body>
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title" data-title="S²FT: Efficient, Scalable and Generalizable LLM
Fine-tuning by Structured Sparsity">S<sup>2</sup>FT: Efficient, Scalable and Generalizable LLM
Fine-tuning by Structured Sparsity</h1>
<div class="is-size-5 publication-authors">
<!-- Paper authors -->
<span class="author-block">
<a href="https://xinyuyang.me/" target="_blank">Xinyu Yang</a><sup>1</sup>,</span>
<span class="author-block">
<a href="https://jixuanleng.com/" target="_blank">Jixuan Leng</a><sup>1</sup>,</span>
<span class="author-block">
<a href="https://guochry.github.io/" target="_blank">Geyang Guo</a><sup>2</sup>
</span>
<span class="author-block">
<a href="https://jiawei-zhao.netlify.app/" target="_blank">Jiawei Zhao</a><sup>3</sup>,
</span>
<span class="author-block">
<a href="https://scholar.google.com/citations?user=80VAi_EAAAAJ&hl=ja" target="_blank">Ryumei
Nakada</a><sup>4</sup>,
</span><br>
<span class="author-block">
<a href="https://linjunz.github.io/index.html" target="_blank">Linjun Zhang</a><sup>4</sup>,
</span>
<span class="author-block">
<a href="https://www.huaxiuyao.io/" target="_blank">Huaxiu Yao</a><sup>5</sup>,
</span>
<span class="author-block">
<a href="https://www.andrew.cmu.edu/user/beidic/" target="_blank">Beidi Chen</a><sup>1</sup>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="affliation"><small><sup>1</sup>Carnegie Mellon University <sup>2</sup>Georgia Tech
<sup>3</sup>Caltech <sup>4</sup>Rutgers <sup>5</sup>UNC-Chapel Hill</small></span>
<span class="eql-cntrb"><small><br>xinyuya2, [email protected]</small></span>
</div>
<div class="column has-text-centered">
<!-- Github link -->
<span class="link-block">
<a href="https://github.com/Infini-AI-Lab/S2FT.git" 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>
<!-- ArXiv abstract Link -->
<span class="link-block">
<a href="https://arxiv.org/abs/2412.06289" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
<!-- Poster Link -->
<span class="link-block">
<a href="https://drive.google.com/file/d/1JJBLlhuKbuBgQ3UK7rsfcIUceaTxQmcG/view?usp=sharing" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-image"></i>
</span>
<span>Poster</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Paper abstract -->
<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">Introduction</h2>
<div class="content has-text-justified">
<p>
We introduce <u>S</u>tructured <u>S</u>parse
<u>F</u>ine-<u>T</u>uning (<b>S<sup>2</sup>FT</b>), the first PEFT method for LLMs that achieves high
quality, efficient training, and scalable serving simultaneously.
S<sup>2</sup>FT accomplishes this by <b>“selecting sparsely and computing densely”</b>. It selects a few
heads and channels in the MHA and FFN modules for each Transformer block,
respectively. Next, it co-permutes weight matrices on both sides of the coupled
structures in LLMs to connect the selected components in each layer into a dense
submatrix. Finally, S<sup>2</sup>FT performs in-place gradient updates on all submatrices.
Through theoretical analysis and empirical results, our method prevents overfitting
and forgetting, delivers SOTA performance on both commonsense and arithmetic
reasoning with <b>4.6%</b> and <b>1.3%</b> average improvements compared to LoRA, and
surpasses full FT by <b>11.5%</b> when generalizing to various domains after instruction
tuning. Using our partial backpropagation algorithm, S<sup>2</sup>FT saves training memory
up to <b>3×</b> and improves latency by <b>1.5-2.7×</b> compared to full FT, while delivering
about <b>10%</b> improvement over LoRA on both metrics. We further demonstrate
that the weight updates in S<sup>2</sup>FT can be decoupled into adapters, enabling effective
fusion, fast switch, and efficient parallelism when serving multiple fine-tuned LLMs.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- End paper abstract -->
<!-- Solutions -->
<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">Why S<sup>2</sup>FT ?</h2>
<div class="content has-text-justified" style="border-bottom: none;">
<style>
.table-container {
width: 100%;
max-width: 800px;
overflow-x: auto; /* Adds horizontal scroll if needed */
}
table {
width: 100%;
max-width: 800px;
border-collapse: collapse;
text-align: center;
margin-top: 20px;
border: none;
table-layout: fixed;
/* Ensures no outer border */
}
th {
padding: 10px;
position: relative;
border: none;
}
/* Override any potential border between rows */
tr {
border-top: none;
border-bottom: none;
}
/* Target specific rows */
tr:first-child th,
tr:nth-child(2) th {
border-top: none;
border-bottom: none;
}
.line-container {
display: flex;
flex-direction: column;
gap: 5px;
/* Space between lines */
margin-top: 10px;
}
.line {
width: 100%;
height: 1px;
background-color: #000;
/* Black lines */
}
</style>
<table style="border-collapse: collapse; width: 100%; text-align: center;">
<tr>
<th scope="col"></th>
<th colspan="2">
High Quality
<div class="line-container">
<div class="line" style="height: 2px;"></div>
</div>
</th>
<th colspan="2">
Efficient Training
<div class="line-container">
<div class="line" style="height: 2px;"></div>
</div>
</th>
<th colspan="3">
Scalable Serving
<div class="line-container">
<div class="line" style="height: 2px;"></div>
</div>
</th>
</tr>
<tr>
<th scope="col"></th>
<th scope="col">ID</th>
<th scope="col">OOD</th>
<th scope="col">Time</th>
<th scope="col">Memory</th>
<th scope="col">Fusion</th>
<th scope="col">Switch</th>
<th scope="col">Parallelism</th>
</tr>
<tr>
<th scope="col">Full FT</th>
<th scope="col">✔✔</th>
<th scope="col">✔</th>
<th scope="col">✘</th>
<th scope="col">✘</th>
<th scope="col">✘</th>
<th scope="col">✘</th>
<th scope="col">✘</th>
</tr>
<tr>
<th scope="col">LoRA</th>
<th scope="col">✔</th>
<th scope="col">✘</th>
<th scope="col">✔</th>
<th scope="col">✔</th>
<th scope="col">✔</th>
<th scope="col">✔</th>
<th scope="col">✔</th>
</tr>
<tr>
<th scope="col"><b>S<sup>2</sup>FT</b></th>
<th scope="col">✔</th>
<th scope="col">✔✔</th>
<th scope="col">✔✔</th>
<th scope="col">✔✔</th>
<th scope="col">✔✔</th>
<th scope="col">✔✔</th>
<th scope="col">✔✔</th>
</tr>
</table>
<p>
Compared to LoRA, S<sup>2</sup>FT offers several key advantages: <b>(i) improved OOD performance, (2) enhanced training efficiency (time & memory), (3) better serving scalability (adapter fusion/switch/parallelism)</b>. These features are particularly valuable in real-world PEFT scenarios, where the goal is to effectively combine knowledge from various domains with the base model's capabilities using limited resource.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- End Solutions -->
<!-- Why -->
<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">Observation</h2>
<div class="content has-text-justified">
<p>
When generalizing to complex reasoning tasks, the performance ranking emerges as: Sparse Fine-tuning (SpFT) > Full FT > LoRA. SpFT effectively transfers reasoning abilities to commonsense domains, while LoRA exhibits significant performance drops in far OOD generalization. This indicates (i) freezing a larger fraction of base model parameters can retain more pre-trained abilities, and (ii) approximating high-dimensional gradients with low-rank decomposition may overfit fine-tuning data and hinder the model from generalization. Since LLMs are typically pre-trained on high-quality data, SpFT emerges as the preferred choice for fine-tuning on domain-specific data of varying quality.
</p>
</div>
<div class="figure">
<img src="static/images/observation.jpg" alt="Robustness and Scalability" width="800" height="400" />
</div>
<div class="content has-text-justified">
<p>
Our findings are further supported by a counterintuitive observation when selecting trainable channels in S²FT using different metrics (weight, activation, and gradients). Surprisingly, selecting channels with the smallest activations leads to improved performance, while selecting those with the largest activations/gradients degrades it. This suggests that for LLMs, it is crucial to preserve more task-relevant advanced skills during pre-training when injecting knowledge from fine-tuning data.
</p>
</div>
<table style="border-collapse: collapse; width: 100%; text-align: center;">
<tr style="font-weight: bold;">
<th style="padding: 8px; text-align: left; border-bottom: 2px solid black; border-top: 2px solid black;" rowspan="2">Task</th>
<th style="padding: 8px; border-bottom: 1px solid black; border-top: 2px solid black;">S²FT-R</th>
<th style="padding: 8px; border-bottom: 1px solid black; border-top: 2px solid black;" colspan="2">S²FT-W</th>
<th style="padding: 8px; border-bottom: 1px solid black; border-top: 2px solid black;" colspan="2">S²FT-A</th>
<th style="padding: 8px; border-bottom: 1px solid black; border-top: 2px solid black;" colspan="2">S²FT-S</th>
<th style="padding: 8px; border-bottom: 1px solid black; border-top: 2px solid black;" colspan="2">S²FT-G</th>
</tr>
<tr style="font-weight: bold;">
<th style="padding: 8px; border-bottom: 2px solid black;"></th>
<th style="padding: 8px; border-bottom: 2px solid black;">Large</th>
<th style="padding: 8px; border-bottom: 2px solid black;">Small</th>
<th style="padding: 8px; border-bottom: 2px solid black;">Large</th>
<th style="padding: 8px; border-bottom: 2px solid black;">Small</th>
<th style="padding: 8px; border-bottom: 2px solid black;">Large</th>
<th style="padding: 8px; border-bottom: 2px solid black;">Small</th>
<th style="padding: 8px; border-bottom: 2px solid black;">Large</th>
<th style="padding: 8px; border-bottom: 2px solid black;">Small</th>
</tr>
<tr style="font-weight: bold;">
<td style="padding: 8px; text-align: left; border-bottom: 1px solid #eee; width: 120px;">Knowledge</td>
<td style="padding: 8px; border-bottom: 1px solid #eee; min-width: 80px;">86.6</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">85.9<sub style="color: red;">(-0.7)</sub></td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">85.3<sub style="color: red;">(-1.3)</sub></td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">84.7<sub style="color: red;">(-1.9)</sub></td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">87.3<sub style="color: green;">(+0.7)</sub></td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">85.1<sub style="color: red;">(-1.5)</sub></td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">87.2<sub style="color: green;">(+0.6)</sub></td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">85.4<sub style="color: red;">(-1.2)</sub></td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">86.2<sub style="color: red;">(-0.4)</sub></td>
</tr>
<tr style="font-weight: bold;">
<td style="padding: 8px; text-align: left; border-bottom: 2px solid black; width: 120px;">Arithmetic</td>
<td style="padding: 8px; border-bottom: 2px solid black; min-width: 80px;">79.6</td>
<td style="padding: 8px; border-bottom: 2px solid black;">78.4<sub style="color: red;">(-1.2)</sub></td>
<td style="padding: 8px; border-bottom: 2px solid black;">78.4<sub style="color: red;">(-1.2)</sub></td>
<td style="padding: 8px; border-bottom: 2px solid black;">77.1<sub style="color: red;">(-2.5)</sub></td>
<td style="padding: 8px; border-bottom: 2px solid black;">80.0<sub style="color: green;">(+0.4)</sub></td>
<td style="padding: 8px; border-bottom: 2px solid black;">76.8<sub style="color: red;">(-2.8)</sub></td>
<td style="padding: 8px; border-bottom: 2px solid black;">79.8<sub style="color: green;">(+0.2)</sub></td>
<td style="padding: 8px; border-bottom: 2px solid black;">77.8<sub style="color: red;">(-1.8)</sub></td>
<td style="padding: 8px; border-bottom: 2px solid black;">79.5<sub style="color: red;">(-0.1)</sub></td>
</tr>
</table>
</div>
</div>
</section>
<!-- End Why -->
<!-- Results -->
<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">S<sup>2</sup>FT and Results</h2>
<div class="content has-text-justified">
<p class="has-text-left is-size-4">
<strong>Method Description</strong>
</p>
<div class="figure">
<img src="static/images/introduction.jpg" alt="S<sup>2</sup>FT"/>
</div>
<p>
First, we sparsely select a few attention heads and channels within the coupled structures of the MHA and FFN modules (see the definition of coupled structures in our paper) as the trainable parameters. Next, we co-permute the weight matrices on both sides of these structures, enabling dense gradient computation only for the selected components. While we demonstrate S<sup>2</sup>FT by selecting the same heads/channels on both sides for clarity, our approach also supports asymmetric selection strategies.
</p>
</div>
<div class="content has-text-justified">
<p class="has-text-left is-size-4">
<strong>Results: High Quality</strong>
</p>
<p>
S<sup>2</sup>FT outperforms Full FT, LoRA, and DoRA on both commonsense and arithmetic reasoning tasks.
</p>
<div style="width: 100%; max-width: 1000px;">
<table style="border-collapse: collapse; width: 100%; text-align: center; margin-bottom: 40px;">
<tr>
<td colspan="11" style="padding: 12px; text-align: center; border-bottom: 1px solid #eee; font-weight: bold;">Fine-tuning LLaMA-3-8B on Commonsense Reasoning Tasks</td>
</tr>
<tr style="font-weight: bold;">
<th style="padding: 8px; text-align: left; border-bottom: 2px solid black; width: 11%;">Method</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 9%;">#Param</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">BoolQ</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">PIQA</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">SIQA</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">HellaSwag</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;"> Wino</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">ARC-e</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">ARC-c</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">OBQA</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">Avg. ↑</th>
</tr>
<!-- First table rows... -->
<tr style="font-weight: bold;">
<td style="padding: 8px; text-align: left; border-bottom: 1px solid #eee;">Full FT</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">100</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">73.9</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">86.2</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">79.1</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">93.1</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">85.8</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">88.1</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">78.2</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">84.0</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">83.6</td>
</tr>
<tr style="font-weight: bold;">
<td style="padding: 8px; text-align: left; border-bottom: 1px solid #eee;">LoRA</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">0.70</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">70.8</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">85.2</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">79.7</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">92.5</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">84.9</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">88.9</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">78.7</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">84.4</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">82.5</td>
</tr>
<tr style="font-weight: bold;">
<td style="padding: 8px; text-align: left; border-bottom: 1px solid #eee;">DoRA</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">0.71</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">74.6</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">89.3</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">79.9</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">95.5</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">85.6</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">90.5</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">80.4</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">85.8</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">85.2</td>
</tr>
<tr style="background-color: #e6f3ff; font-weight: bold;">
<td style="padding: 8px; text-align: left; border-bottom: 1px solid #eee; font-weight: bold;">S<sup>2</sup>FT</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">0.70</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">75.0</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">89.0</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">80.7</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">96.5</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">88.0</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">92.5</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">83.4</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">87.8</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">86.6</td>
</tr>
</table>
<table style="border-collapse: collapse; width: 100%; text-align: center;">
<tr>
<td colspan="10" style="padding: 12px; text-align: center; border-bottom: 1px solid #eee; font-weight: bold;">Fine-tuning LLaMA-3-8B on Arithmetic Reasoning Tasks</td>
</tr>
<tr style="font-weight: bold;">
<th style="padding: 8px; text-align: left; border-bottom: 2px solid black; width: 11%;">Method</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 9%;">#Param</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 16%;">MultiArith</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">GSM8K</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">AddSub</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">AQuA</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">SingleEq</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">SVAMP</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">MAWPS</th>
<th style="padding: 8px; border-bottom: 2px solid black; width: 8%;">Avg. ↑</th>
</tr>
<tr style="font-weight: bold;">
<td style="padding: 8px; text-align: left; border-bottom: 1px solid #eee;">Full FT</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">100</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">99.2</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">62.0</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">93.9</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">26.8</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">96.7</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">74.0</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">91.2</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">77.7</td>
</tr>
<tr style="font-weight: bold;">
<td style="padding: 8px; text-align: left; border-bottom: 1px solid #eee;">LoRA</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">0.70</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">99.5</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">61.6</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">92.7</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">25.6</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">96.3</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">73.8</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">90.8</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">77.2</td>
</tr>
<tr style="font-weight: bold;">
<td style="padding: 8px; text-align: left; border-bottom: 1px solid #eee;">DoRA</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">0.71</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">98.8</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">62.7</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">92.2</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">26.8</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">96.9</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">74.0</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">91.2</td>
<td style="padding: 8px; border-bottom: 1px solid #eee;">77.5</td>
</tr>
<tr style="background-color: #e6f3ff; font-weight: bold;">
<td style="padding: 8px; text-align: left; border-bottom: 1px solid #eee; font-weight: bold;">S<sup>2</sup>FT</td>
<td style="padding: 8px; border-bottom: 1px solid #eee; font-weight: bold;">0.70</td>
<td style="padding: 8px; border-bottom: 1px solid #eee; font-weight: bold;">99.7</td>
<td style="padding: 8px; border-bottom: 1px solid #eee; font-weight: bold;">65.8</td>
<td style="padding: 8px; border-bottom: 1px solid #eee; font-weight: bold;">93.7</td>
<td style="padding: 8px; border-bottom: 1px solid #eee; font-weight: bold;">31.5</td>
<td style="padding: 8px; border-bottom: 1px solid #eee; font-weight: bold;">97.8</td>
<td style="padding: 8px; border-bottom: 1px solid #eee; font-weight: bold;">76.0</td>
<td style="padding: 8px; border-bottom: 1px solid #eee; font-weight: bold;">92.4</td>
<td style="padding: 8px; border-bottom: 1px solid #eee; ">79.6</td>
</tr>
</table>
</div>
</div>
<div class="content has-text-justified">
<p class="has-text-left is-size-4">
<strong>Results: Efficient Training</strong>
<p>
S<sup>2</sup>FT delivers an average 10% improvement over LoRA on both training memory and time.
</p>
<div class="figure">
<img src="static/images/training_efficiency.jpg" alt="training efficiency" width="900" height="400" />
</div>
</p>
</div>
<div class="content has-text-justified">
<p class="has-text-left is-size-4">
<strong>Results: Scalable Inference</strong>
</p>
<p>
S<sup>2</sup>FT can modify orthogonal low-rank spaces for different tasks, resulting in effective adpater fusion.
</p>
<table style="border-collapse: collapse; width: 100%; text-align: center;">
<thead>
<tr>
<td colspan="7" style="padding: 12px; text-align: center; border-bottom: 1px solid #eee; font-weight: bold;">Fusing Commonsense and Arithmetic Adapters for LLaMA-3-8B</td>
</tr>
<tr style="font-weight: bold;">
<th style="padding: 8px; text-align: left; border-top: 2px solid black; border-bottom: 2px solid black;" rowspan="2">Task</th>
<th style="padding: 8px; border-top: 2px solid black; border-bottom: 1px solid black;" colspan="3">LoRA</th>
<th style="padding: 8px; border-top: 2px solid black; border-bottom: 1px solid black;" colspan="3">S²FT</th>
</tr>
<tr style="font-weight: bold;">
<th style="padding: 8px; border-bottom: 2px solid black;">Adapter 1</th>
<th style="padding: 8px; border-bottom: 2px solid black;">Adapter 2</th>
<th style="padding: 8px; border-bottom: 2px solid black;">Fused</th>
<th style="padding: 8px; border-bottom: 2px solid black;">Adapter 1</th>
<th style="padding: 8px; border-bottom: 2px solid black;">Adapter 2</th>
<th style="padding: 8px; border-bottom: 2px solid black;">Fused</th>
</tr>
</thead>
<tbody>
<tr style="font-weight: bold;">
<td style="padding: 8px; text-align: left; border-bottom: 2px solid black;">Commonsense</td>
<td style="padding: 8px; border-bottom: 2px solid black;">83.1</td>
<td style="padding: 8px; border-bottom: 2px solid black; color: #999;">32.1</td>
<td style="padding: 8px; border-bottom: 2px solid black;">79.8<sub style="color: red;">(-3.3)</sub></td>
<td style="padding: 8px; border-bottom: 2px solid black;">86.6</td>
<td style="padding: 8px; border-bottom: 2px solid black; color: #999;">42.3</td>
<td style="padding: 8px; border-bottom: 2px solid black;">84.0<sub style="color: red;">(-2.6)</sub></td>
</tr>
<tr style="font-weight: bold;">
<td style="padding: 8px; text-align: left; border-bottom: 2px solid black;">Arithmetic</td>
<td style="padding: 8px; border-bottom: 2px solid black; color: #999;">12.0</td>
<td style="padding: 8px; border-bottom: 2px solid black;">77.2</td>
<td style="padding: 8px; border-bottom: 2px solid black;">71.6<sub style="color: red;">(-5.6)</sub></td>
<td style="padding: 8px; border-bottom: 2px solid black; color: #999;">12.8</td>
<td style="padding: 8px; border-bottom: 2px solid black;">79.6</td>
<td style="padding: 8px; border-bottom: 2px solid black;">75.3<sub style="color: red;">(-4.3)</sub></td>
</tr>
</tbody>
</table>
<p>
S<sup>2</sup>FT enables scalable and efficient adapter switch and parallelism by reducing matmul operations.
</p>
<div class="figure">
<img src="static/images/scalable_serving.png" alt="scalable serving" width="900" height="400" />
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Results -->
<!-- Discussion -->
<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">Conclusion and Future Work</h2>
<div class="content has-text-justified">
<p>
This work introduces S<sup>2</sup>FT, a novel PEFT family that is generalizable, efficient, and scalable. Compared to LoRA，S<sup>2</sup>FT improves the generalization ability on downstream tasks while reduce 10% training time and memory. Furthermore, S<sup>2</sup>FT's enables scalable serving of thousands of adapters simultaneously. These comprehensive improvements in quality, efficiency, and scalability make S<sup>2</sup>FT particularly valuable for the large-scale, real-world deployment of foundation models in various domains. Future research directions include exploring the controllability in S<sup>2</sup>FT, which enablee the separation of domain-specific knowledge into distinct parameters. This capability could significantly advance sparse training and inference techniques for LLMs, particularly in MOE architecture design.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- Disucssion -->
<!--BibTex citation -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@inproceedings{yang2024s2ft,
title={S<sup>2</sup>FT: Efficient, Scalable and Generalizable LLM Fine-tuning by Structured Sparsity},
author={Yang, Xinyu and Leng, Jixuan and Guo, Geyang and Zhao, Jiawei and Nakada, Ryumei and Zhang, Linjun and Yao, Huaxiu and Chen, Beidi},
booktitle={The 38th Conference on Neural Information Processing Systems (NeurIPS)},
year={2024}
}</code></pre>
</div>
</section>
<!--End BibTex citation -->
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This page was built using the <a href="https://github.com/eliahuhorwitz/Academic-project-page-template"
target="_blank">Academic Project Page Template</a> which was adopted from the <a
href="https://nerfies.github.io" target="_blank">Nerfies</a> project page.
You are free to borrow the of this website, we just ask that you link back to this page in the footer.
<br> This website is licensed under a <a rel="license"
href="http://creativecommons.org/licenses/by-sa/4.0/" target="_blank">Creative
Commons Attribution-ShareAlike 4.0 International License</a>.
</p>
</div>
</div>
</div>
</div>
</footer>
<!-- Statcounter tracking code -->
<!-- You can add a tracker to track page visits by creating an account at statcounter.com -->
<!-- End of Statcounter Code -->
</body>
</html>