-
Notifications
You must be signed in to change notification settings - Fork 0
/
index.html
528 lines (464 loc) · 26.7 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
<!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="DESCRIPTION META TAG">
<meta property="og:title" content="SOCIAL MEDIA TITLE TAG"/>
<meta property="og:description" content="SOCIAL MEDIA DESCRIPTION TAG TAG"/>
<meta property="og:url" content="URL OF THE WEBSITE"/>
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X630-->
<meta property="og:image" content="static/image/your_banner_image.png" />
<meta property="og:image:width" content="1200"/>
<meta property="og:image:height" content="630"/>
<meta name="twitter:title" content="TWITTER BANNER TITLE META TAG">
<meta name="twitter:description" content="TWITTER BANNER DESCRIPTION META TAG">
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X600-->
<meta name="twitter:image" content="static/images/your_twitter_banner_image.png">
<meta name="twitter:card" content="summary_large_image">
<!-- Keywords for your paper to be indexed by-->
<meta name="keywords" content="KEYWORDS SHOULD BE PLACED HERE">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Accelerating GPT-2 Inference with TVM DLight Auto-Scheduling</title>
<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">Accelerating GPT-2 Inference with TVM DLight Auto-Scheduling</h1>
<div class="is-size-5 publication-authors">
<!-- Paper authors -->
<span class="author-block">
<a href="mailto:[email protected]" target="_blank">Charlie Ruan (cfruan)</a><sup>*</sup>,</span>
<span class="author-block">
<a href="mailto:[email protected]" target="_blank">Xun Zhou (xzhou2)</a><sup>*</sup>,</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block">Carnegie Mellon University<br>15-618 F23 Final Project Proposal</span>
<span class="eql-cntrb"><small><br><sup>*</sup>Indicates Equal Contribution</small></span>
</div>
<div class="column has-text-centered">
<div class="column has-text-centered">
<div class="publication-links">
<!-- Arxiv PDF link -->
<span class="link-block">
<a href="static/pdfs/project_proposal.pdf" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Proposal</span>
</a>
</span>
<span class="link-block">
<a href="static/pdfs/project_milestone.pdf" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Milestone</span>
</a>
</span>
<div class="publication-links">
<!-- Arxiv PDF link -->
<span class="link-block">
<a href="static/pdfs/final_report.pdf" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Final Report</span>
</a>
</span>
<!-- Github link -->
<span class="link-block">
<a href="https://github.com/rickzx/618-project/blob/main/code" 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>
<a href="static/pdfs/poster.pdf">
<img src="static/images/poster.png" alt="Image Description" class="centered-image-pad">
</a>
<!-- Paper abstract -->
<section class="section hero is-info">
<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">
<p>
The goal of this project is two-fold: optimizing GPT-2’s inference performance with TVM on various GPU backends (e.g. CUDA, ROCm, WebGPU), and analyzing and tuning TVM’s auto-scheduler DLight. For the first part, we will implement GPT-2’s inference on the open-source framework MLC-LLM, which allows us to optimize LLM workloads with TVM conveniently. For the second part, we will analyze the effect of various transformations performed by DLight (both performance and code generation), perform benchmarks across devices, and further tune the scheduler to optimize GPT-2’s performance for a specific backend. </p>
</div>
</div>
</div>
</div>
</section>
<!-- End paper abstract -->
<!-- Background -->
<section class="section hero is-small">
<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">Background</h2>
<div class="content has-text-justified">
<p>
GPT-2 is a transformer-based decoder-only language model that can benefit significantly from parallelism due to its computational intensity. This project only focuses on the inference phase and focuses on the parallelism brought by multi-core GPUs, including CUDA (Nvidia GPUs), ROCm (AMD GPUs), and WebGPU (available on Chrome). Specifically, the matrix multiplications in GPT-2 can benefit significantly from such highly parallelizable backends.
</p><p>
To optimize the distribution and execution of the workload, we leverage TVM, a domain-specific framework for machine learning compilation. To use TVM, there are two general steps: defining the task, and optimizing the task with a scheduler. For the first step, we will use the recently developed open-source framework MLC-LLM, which is built on top of TVM and dedicated to large language models like GPT-2. For the second step, we focus on DLight, an auto-scheduler in TVM that optimizes machine learning workloads. An example usage of DLight is shown below.
</p>
<img src="static/images/dlight-example.png" alt="dlight example" width="50%" class="center"/>
<p>
Some general parallelizing optimizations include reducing data transfer under a GPU’s multi-core setting, dividing the workload to best leverage a GPU’s architecture (e.g. structuring of for loops), and incorporating various other techniques such as tiling. With these, we optimize the performance of GPT-2, measured by the number of tokens per second (for both the prefilling and decoding phases).
</p></div>
</div>
</div>
</div>
</section>
<!-- End background -->
<!-- Challenges -->
<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">Challenges</h2>
<div class="content has-text-justified">
<p>
The inference of GPT-2 resembles that of most transformer-based decoder-only language models, revolving around various matrix-matrix multiplications. Given that we use a single device, the main characteristics of the workload include: it is computationally intensive due to the massive number of parameters, and it is easily bounded by memory bandwidth when using techniques like KV cache.
</p>
<p>Some constraints that make the system hard to parallelize include the fact that inference in LLMs like GPT-2 is in an autoregressive manner, meaning that the generation of a token depends on the previous token—so we have to generate one token at a time sequentially. This poses a challenge because we can only try to exploit parallelism at a much lower level, mainly the operators like matrix multiplications.
</p>
<p>
While there is a plethora of optimization tricks and techniques, we focus on operator-level optimization. As mentioned above, we will use DLight to achieve such optimization. Since the project includes domain-specific programming, one challenge for us is to gain domain knowledge (e.g. what are the typical techniques for operator optimization, including how GEMV works), and another is to understand the existing software stack of TVM so that we can tune the scheduler. Since we aim to analyze the effect of various scheduling rules, one more challenge is that we need to understand the semantics of the generated code, some of which are not typical code we use (e.g. the WebGPU shading language).
</p>
</div>
</div>
</div>
</section>
<!-- End challenges -->
<!-- Resources -->
<section class="section hero is-small">
<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">Resources</h2>
<div class="content has-text-justified">
<p>
To test GPT-2’s inference, we will use a single RTX 4090 for the CUDA backend, a single Rx 7900 for the ROCm backend, and an Apple Macbook for the WebGPU backend.
</p>
<p>
The first step of our project is to implement GPT-2 in MLC-LLM (i.e. specifying the algorithm to optimize). We will work with the existing code base <a href="https://github.com/mlc-ai/mlc-llm" target="_blank">MLC-LLM</a> and submit a Pull Request.
</p>
The second step of our project is to analyze and tune the DLight schedule. We will refer to the codebase of DLight found <a href="https://github.com/apache/tvm/tree/unity/python/tvm/dlight" target="_blank">here</a>.
<p>
Some other resources we will refer to include the <a href="https://arxiv.org/abs/1802.04799">TVM paper</a>, the <a href="https://arxiv.org/abs/2311.02103">Relax paper (a new abstraction under TVM)</a>, and perhaps many more.
</p>
For specific languages like WebGPU shader language, we can refer to its <a href="https://www.w3.org/TR/WGSL/">documentation</a>.
</div>
</div>
</div>
</section>
<!-- End resources -->
<!-- Goals and Deliverables -->
<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">Goals and Deliverables</h2>
<div class="content has-text-justified">
<p>This project is mainly an analysis project, hoping to understand how various scheduling rules in DLight affect the performance of LLM inference on various backends. With that being said, we plan to achieve the workload-defining part and the analysis part; we hope to achieve better results than the given scheduling rule on a specific backend by tuning the scheduler. We put it more concretely below.</p>
<h4 class="title is-5">PLAN TO ACHIEVE</h4>
<ul>
<li>Implement GPT-2 inference in MLC-LLM, so that it can be optimized with the default DLight schedule.</li>
<li>Benchmark GPT-2 inference performance and speedup across various backends.</li>
<li>For each scheduling rule provided by DLight:
<ul>
<li>Analyze how it affects: the target code generated (e.g. CUDA code), and the performance gained.</li>
<li>Compare how much speedup each backend gains for each rule.</li>
<li>If applicable, within a specific scheduling rule, analyze how hyperparameters affect performance on different backends</li>
</ul>
</li>
</ul>
<h4 class="title is-5">HOPE TO ACHIEVE</h4>
<p>Pick a specific backend, say WebGPU, and tune the scheduling rule tailored towards that backend, hoping to achieve better performance than the default schedule used.</p>
<h4 class="title is-5">DEMO PLAN</h4>
<p>For the poster session, we will dedicate a portion to give an overview of the GPT-2 inference workload, and how it can be optimized via TVM (the task-defining portion, and the performance-optimizing portion). Afterward, for each scheduling rule, we will describe what it does, give an example change in the code (e.g. after this transformation, how this CUDA kernel changed), and draw a graph that compares how each backend benefits from this rule. If we achieve our stretch goal, we will add another section illustrating how we achieve better speedup than the default scheduling rule.</p>
</div>
</div>
</div>
</section>
<!-- End goals -->
<!-- Platform Choice -->
<section class="section hero is-small">
<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">Platform Choice</h2>
<div class="content has-text-justified">
<p>
First of all, we use GPUs because this is the typical go-to type of device for machine learning workloads due to how good they are for matrix operations. We include CUDA (i.e. Nvidia GPU) mostly as a reference since this is the most conventional GPU backend. We include WebGPU because while the default schedule may be highly optimized for common backends like CUDA, it may not be the case for less conventional ones like WebGPU. This gives us the space to potentially gain more speedup by tuning the scheduling rule. We include ROCm (i.e. AMD GPU) as a middle ground.
</p>
</div>
</div>
</div>
</section>
<!-- End platform -->
<!-- Preliminary 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">Preliminary Results: Benchmarks across platforms</h2>
<div class="content has-text-justified">
<p>After implementing GPT 2, we ran its inference with TVM on the CUDA backend (with a single RTX 4090), on a Macbook Pro (Intel), and on the WebGPU backend (on the same Macbook Pro). We use the number of tokens per second to describe the performance. Specifically, we prefill 322 tokens (the entire Gettysburg Address) and let the model generate 64 tokens autoregressively.</p>
<table border="1">
<tr>
<th>Prefill (tokens/sec)</th>
<th>Decode (tokens/sec)</th>
</tr>
<tr>
<td>RTX 4090</td>
<td>4419.0</td>
<td>102.9</td>
</tr>
<tr>
<td>MacBook Pro (Intel)</td>
<td>28.9</td>
<td>1.1</td>
</tr>
<tr>
<td>WebGPU (same MacBook Pro)</td>
<td>10.96</td>
<td>0.81</td>
</tr>
</table>
<p>From the result, one thing to notice is that, despite leveraging the same hardware, we are receiving worse results on the WebGPU backend. This makes one of the goal for us to explore better operator-level optimization for the WebGPU backend to utilize its hardware better, hopefully closing the gap. Note that we mentioned in the proposal that we will also be testing on the ROCm backend (AMD GPU) – due to hardware issues, we will supplement the result in the final submission.</p>
</div>
</div>
</div>
</section>
<section class="section hero is-small">
<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">Preliminary Results: Effects of each operator-level optimization</h2>
<div class="content has-text-justified">
<p>Recall that, as mentioned in the proposal, we focus on operator-level optimization within TVM, and such optimization is done in a sub-framework called DLight. The usage of DLight can be illustrated as below:</p>
<p>After implementing GPT-2 in MLC LLM, another benchmarking we did was to determine how much effect each rule has on the performance. Below are the results after running the same experiment on RTX 4090 (prefilling 322 tokens, generating 64 tokens autoregressively):</p>
<table border="1">
<tr>
<th>Schedule</th>
<th>Prefill (tokens/sec)</th>
<th>Prefill Speedup</th>
<th>Decode (tokens/sec)</th>
<th>Decode Speedup</th>
</tr>
<tr>
<td>Baseline</td>
<td>70.0</td>
<td>1x</td>
<td>3.5</td>
<td>1x</td>
</tr>
<tr>
<td>Matmul</td>
<td>3321.4</td>
<td>47.4x</td>
<td>3.5</td>
<td>1x</td>
</tr>
<tr>
<td>GEMV</td>
<td>70.1</td>
<td>1x</td>
<td>53.5</td>
<td>15.3x</td>
</tr>
<tr>
<td>Reduction</td>
<td>70.1</td>
<td>1x</td>
<td>60.7</td>
<td>17.3x</td>
</tr>
<tr>
<td>Full</td>
<td>4419.0</td>
<td>63.1x</td>
<td>102.9</td>
<td>29.4x</td>
</tr>
</table>
<p>“Baseline” denotes without using any optimization, and “full” denotes using all optimization. Note that we consider Reduction() and GeneralReduction() as a single rule; we also note that due to the current implementation, we need to include Fallback() in all schedules, hence considering it as part of the baseline.</p>
<p>We observe that matrix multiplication contributes the most speedup for the prefill stage, which makes sense because during prefill, most operations are matrix-matrix multiplication. This is due to how we have multiple values in the Q matrix. Meanwhile, GEMV, which optimizes matrix-vector multiplication, contributes to the speedup during decoding. Similarly, this makes sense because we only have a single query token in the Q matrix during the decoding stage.</p>
</div>
</div>
</div>
</section>
<!-- Final Results -->
<section class="section hero is-small">
<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">Further Optimization: Tuning GEMV</h2>
<div class="content has-text-justified">
<img src="static/images/seq_gemv.png" alt="Sequential GEMV" class="centered-image"/>
<p>To parallelize GEMV, TVM does the following steps:</p>
<ul>
<li>Break the N rows into groups of TS contiguous rows; each threadBlock takes a group.
<p>(Note each threadBlock’s work is independent from each other since we split across the spatial axis)</p>
</li>
<li>Each block uses a thread layout of TS by TR
<p>(i.e. each block has TS*TR threads; as a result, each thread has TS*K/(TS*TR) = K/TR amount of work)</p>
</li>
<li>Each thread block works on a tile of shape TILE_S by TILE_R at a time
<p>(Again, TILE_S is split across the spatial axis, and TILE_R across the reduction)</p>
</li>
</ul>
<img src="static/images/par_gemv.png" alt="Parallel GEMV" class="centered-image"/>
</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">Further Optimization: Final Result</h2>
<div class="content has-text-justified">
<img src="static/images/result.png" alt="Final result" class="centered-image"/>
</div>
</div>
</div>
</section>
<!-- Schedule -->
<section class="section hero is-small">
<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">Schedule</h2>
<div class="content has-text-justified">
<table border="1">
<tr>
<th class="schedule-column">SCHEDULE</th>
<th>Activities</th>
</tr>
<tr>
<td>Week 1<br>(11/13 - 11/19)</td>
<td>
<ul>
<li>Brainstorm project scope and submit proposal</li>
<li>Start implementing GPT 2 in MLC-LLM</li>
</ul>
</td>
</tr>
<tr>
<td>Week 2<br>(11/20 - 11/26)</td>
<td>
<ul>
<li>Finish and verify GPT 2 implementation in MLC-LLM
</li></ul>
</td>
</tr>
<tr>
<td>Week 3<br>(11/27 - 12/3)</td>
<td>
<ul>
<li>Benchmark speedup for different backends with default scheduler</li>
<li>Analyze effects of each scheduling rule for each rule</li>
<li>Complete milestone report (due 12/2)</li>
</ul>
</td>
</tr>
<tr>
<td>Week 4<br>(12/4 - 12/10)</td>
<td>
<ul>
<li>Learn how each scheduler rule transforms the program</li>
<li>Learn mechanisms of tuning the DLight scheduler (through machine learning, grid search, etc.); locate the hyperparameters</li>
</ul>
</td>
</tr>
<tr>
<td>Week 5<br>(12/11 - 12/17)</td>
<td>
<ul>
<li>Tune scheduler to achieve a better speedup for a specific backend</li>
<li>Complete final report and poster (due 12/14)</li>
</ul>
</td>
</tr>
</table>
</div>
</div>
</div>
</section>
<!-- End schedule -->
<!--BibTex citation -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">Reference</h2>
<pre><code>
[1] T. Chen, T. Moreau, Z. Jiang, L. Zheng, E. Yan, M. Cowan, H. Shen, L. Wang, Y. Hu, L.
Ceze, C. Guestrin, and A. Krishnamurthy, Tvm: An automated end-to-end optimizing compiler for deep learning (2018), arXiv:1802.04799 [cs.LG].
[2] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, et al., Language models are
unsupervised multitask learners, OpenAI blog 1, 9 (2019)
[3] MLC Team, MLC-LLM (2023) https://github.com/mlc-ai/mlc-llm
[4] U Jamil, LLaMA Explained (2023) https://github.com/hkproj/pytorch-llama-notes
[5] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N.
Gimelshein, L. Antiga, A. Desmaison, A. K ̈opf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, Pytorch: An imperative style, high-performance deep learning library (2019), arXiv:1912.01703 [cs.LG].
[6] J. Thickstun, D. Hall, C. Donahue, and P. Liang, Anticipatory music transformer (2023),
arXiv:2306.08620 [cs.SD].
[7] MLC Team, Dlight-Bench (2023) https://github.com/mlc-ai/dlight-bench
</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>