-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathpytorch.html
More file actions
538 lines (523 loc) · 40.3 KB
/
Copy pathpytorch.html
File metadata and controls
538 lines (523 loc) · 40.3 KB
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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>PyTorch in One Hour — Visual Map</title>
<link href="https://fonts.googleapis.com/css2?family=Lora:ital,wght@0,400;0,600;0,700;1,400&family=JetBrains+Mono:wght@300;400;500;700&family=Libre+Baskerville:ital,wght@0,400;0,700;1,400&display=swap" rel="stylesheet">
<style>
:root {
--paper:#f9f6f0; --paper2:#f2ede4; --ink:#1a1810; --ink2:#3a3428; --ink3:#6b6155;
--torch:#ee4c2c; --torch2:#c73e23; --torch-bg:#fff0ed;
--blue:#1a4480; --blue-bg:#edf3fc; --green:#1a6640; --green-bg:#edf7f1;
--gold:#8a6a00; --gold-bg:#fdf8e8; --rule:#d5cec4;
--serif:'Lora',Georgia,serif; --mono:'JetBrains Mono','Courier New',monospace;
--bask:'Libre Baskerville',Georgia,serif;
}
*{box-sizing:border-box;margin:0;padding:0;}
body{background:var(--paper);color:var(--ink);font-family:var(--serif);min-height:100vh;position:relative;}
body::before{content:'';position:fixed;inset:0;background-image:url("data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' width='400' height='400'%3E%3Cfilter id='n'%3E%3CfeTurbulence type='fractalNoise' baseFrequency='0.75' numOctaves='4' stitchTiles='stitch'/%3E%3C/filter%3E%3Crect width='100%25' height='100%25' filter='url(%23n)' opacity='0.03'/%3E%3C/svg%3E");pointer-events:none;z-index:0;}
body::after{content:'';position:fixed;inset:0;background-image:repeating-linear-gradient(transparent,transparent 27px,rgba(180,170,155,0.18) 27px,rgba(180,170,155,0.18) 28px);pointer-events:none;z-index:0;}
.page{position:relative;z-index:1;max-width:960px;margin:0 auto;padding:56px 48px 80px;}
.back-link{display:inline-flex;align-items:center;gap:8px;font-family:var(--mono);font-size:9px;letter-spacing:.18em;text-transform:uppercase;color:var(--ink3);text-decoration:none;margin-bottom:36px;transition:color .2s;}
.back-link:hover{color:var(--torch);}
.header{margin-bottom:56px;padding-bottom:28px;border-bottom:2px solid var(--ink);position:relative;}
.header::after{content:'';position:absolute;bottom:-5px;left:0;width:100%;height:1px;background:var(--rule);}
.header-kicker{font-family:var(--mono);font-size:10px;letter-spacing:.25em;text-transform:uppercase;color:var(--torch);margin-bottom:10px;}
h1{font-family:var(--bask);font-size:clamp(28px,4vw,48px);font-weight:700;line-height:1.1;color:var(--ink);margin-bottom:8px;}
h1 em{font-style:italic;color:var(--torch);}
.header-sub{font-family:var(--serif);font-size:15px;color:var(--ink3);font-style:italic;max-width:600px;line-height:1.6;}
.header-meta{display:flex;gap:20px;margin-top:16px;flex-wrap:wrap;}
.tag{font-family:var(--mono);font-size:9px;letter-spacing:.12em;text-transform:uppercase;border:1px solid var(--rule);padding:3px 10px;color:var(--ink3);background:var(--paper2);}
.sec-header{display:flex;align-items:baseline;gap:16px;margin-bottom:20px;margin-top:48px;}
.sec-num{font-family:var(--mono);font-size:11px;color:var(--torch);letter-spacing:.1em;flex-shrink:0;}
.sec-title{font-family:var(--bask);font-size:18px;font-weight:700;color:var(--ink);border-bottom:1px solid var(--rule);padding-bottom:6px;flex:1;}
.three-cols{display:grid;grid-template-columns:repeat(3,1fr);gap:2px;background:var(--ink);border:2px solid var(--ink);margin-bottom:8px;}
.comp-card{background:var(--paper);padding:24px 20px;position:relative;overflow:hidden;transition:background .2s;}
.comp-card:hover{background:var(--paper2);}
.comp-card::before{content:'';position:absolute;top:0;left:0;right:0;height:3px;}
.comp-card.tensor::before{background:var(--torch);}
.comp-card.autograd::before{background:var(--blue);}
.comp-card.dl::before{background:var(--green);}
.comp-num{font-family:var(--mono);font-size:9px;color:var(--ink3);letter-spacing:.15em;margin-bottom:6px;}
.comp-name{font-family:var(--bask);font-size:16px;font-weight:700;margin-bottom:4px;}
.comp-card.tensor .comp-name{color:var(--torch);}
.comp-card.autograd .comp-name{color:var(--blue);}
.comp-card.dl .comp-name{color:var(--green);}
.comp-sub{font-family:var(--mono);font-size:9px;color:var(--ink3);margin-bottom:10px;letter-spacing:.05em;}
.comp-desc{font-size:12px;line-height:1.65;color:var(--ink2);}
.comp-analogy{margin-top:10px;font-size:11px;font-style:italic;color:var(--ink3);border-top:1px solid var(--rule);padding-top:8px;}
.tensor-grid{display:grid;grid-template-columns:repeat(4,1fr);gap:2px;background:var(--rule);border:1px solid var(--rule);margin-bottom:16px;}
.tensor-cell{background:var(--paper);padding:18px 14px;text-align:center;transition:background .2s;}
.tensor-cell:hover{background:var(--torch-bg);}
.tensor-rank{font-family:var(--mono);font-size:9px;color:var(--ink3);letter-spacing:.1em;margin-bottom:8px;}
.tensor-name{font-family:var(--bask);font-size:14px;font-weight:700;color:var(--ink);margin-bottom:8px;}
.tensor-vis{height:52px;display:flex;align-items:center;justify-content:center;margin-bottom:8px;}
.tensor-code{font-family:var(--mono);font-size:9px;color:var(--torch);letter-spacing:.03em;line-height:1.5;}
.tensor-example{font-size:11px;color:var(--ink3);margin-top:6px;font-style:italic;}
.scalar-dot{width:16px;height:16px;border-radius:50%;background:var(--torch);display:inline-block;}
.vec-bar{display:flex;flex-direction:column;gap:3px;}
.vec-bar span{width:24px;height:8px;background:var(--torch);display:block;opacity:0.7;}
.vec-bar span:first-child{opacity:1;}
.mat-grid{display:grid;grid-template-columns:repeat(3,1fr);gap:2px;width:54px;}
.mat-grid span{height:14px;background:var(--torch);display:block;}
.mat-grid span:nth-child(even){opacity:.5;}
.mat-grid span:nth-child(3n){opacity:.3;}
.tensor3d-wrap{position:relative;width:54px;height:48px;}
.tensor3d-layer{position:absolute;width:34px;height:28px;background:var(--torch);border:1px solid var(--paper);}
.tensor3d-layer:nth-child(1){top:0;left:0;opacity:.3;}
.tensor3d-layer:nth-child(2){top:8px;left:8px;opacity:.6;}
.tensor3d-layer:nth-child(3){top:16px;left:16px;opacity:1;}
.dtype-table{width:100%;border-collapse:collapse;font-family:var(--mono);font-size:11px;border:1px solid var(--rule);}
.dtype-table th{background:var(--ink);color:var(--paper);padding:8px 12px;text-align:left;letter-spacing:.08em;font-size:9px;text-transform:uppercase;}
.dtype-table td{padding:7px 12px;border-bottom:1px solid var(--rule);vertical-align:top;}
.dtype-table tr:nth-child(even) td{background:var(--paper2);}
.dtype-table tr:hover td{background:var(--torch-bg);}
.dtype-key{color:var(--torch);font-weight:700;}
.compgraph-wrap{display:grid;grid-template-columns:1fr 1fr;gap:2px;background:var(--rule);border:1px solid var(--rule);}
.compgraph-panel{background:var(--paper);padding:24px;}
.compgraph-panel h3{font-family:var(--mono);font-size:10px;letter-spacing:.15em;text-transform:uppercase;color:var(--ink3);margin-bottom:16px;padding-bottom:8px;border-bottom:1px solid var(--rule);}
.fwd-flow,.bwd-flow{display:flex;flex-direction:column;gap:2px;}
.fwd-node,.bwd-node{display:flex;align-items:center;gap:12px;padding:10px 14px;background:var(--paper2);border:1px solid var(--rule);transition:background .2s;}
.fwd-node:hover{background:var(--blue-bg);}
.bwd-node:hover{background:var(--gold-bg);}
.fwd-symbol{font-family:var(--mono);font-size:13px;font-weight:700;color:var(--blue);width:40px;text-align:center;flex-shrink:0;}
.bwd-symbol{font-family:var(--mono);font-size:11px;font-weight:700;color:var(--gold);width:48px;text-align:center;flex-shrink:0;}
.fwd-label,.bwd-label{flex:1;}
.fwd-label strong,.bwd-label strong{display:block;font-family:var(--mono);font-size:11px;color:var(--ink);margin-bottom:2px;}
.fwd-label span,.bwd-label span{font-size:10px;color:var(--ink3);font-style:italic;}
.fwd-code{font-family:var(--mono);font-size:9px;color:var(--torch);flex-shrink:0;text-align:right;}
.bwd-code{font-family:var(--mono);font-size:9px;color:var(--gold);flex-shrink:0;text-align:right;}
.chain-rule-box{margin-top:14px;background:var(--gold-bg);border:1px solid #e8d88a;padding:12px 14px;}
.chain-rule-box .label{font-family:var(--mono);font-size:9px;letter-spacing:.12em;text-transform:uppercase;color:var(--gold);margin-bottom:6px;}
.chain-rule-math{font-family:var(--mono);font-size:12px;color:var(--ink);line-height:2;}
.chain-rule-note{font-size:11px;color:var(--ink3);font-style:italic;margin-top:6px;}
.pipeline{display:flex;flex-direction:column;gap:0;border:1px solid var(--ink);}
.pipe-step{display:grid;grid-template-columns:52px 1fr 1fr;gap:0;border-bottom:1px solid var(--rule);min-height:80px;}
.pipe-step:last-child{border-bottom:none;}
.pipe-num{background:var(--ink);color:var(--paper);display:flex;flex-direction:column;align-items:center;justify-content:center;font-family:var(--mono);font-size:18px;font-weight:700;gap:2px;}
.pipe-num small{font-size:7px;letter-spacing:.1em;text-transform:uppercase;opacity:.5;}
.pipe-info{padding:16px 18px;border-right:1px solid var(--rule);}
.pipe-name{font-family:var(--bask);font-size:15px;font-weight:700;color:var(--ink);margin-bottom:4px;}
.pipe-desc{font-size:12px;color:var(--ink3);line-height:1.6;}
.pipe-gotcha{margin-top:8px;font-size:11px;background:var(--torch-bg);border-left:3px solid var(--torch);padding:5px 8px;color:var(--torch2);font-style:italic;}
.pipe-code{padding:14px 16px;background:var(--paper2);font-family:var(--mono);font-size:10px;line-height:1.7;color:var(--ink);overflow:hidden;}
.pipe-code .comment{color:var(--ink3);font-style:italic;}
.pipe-code .kw{color:var(--blue);font-weight:700;}
.pipe-code .fn{color:var(--torch);}
.pipe-code .str{color:var(--green);}
.pipe-code .num{color:var(--gold);}
.loop-box{border:2px solid var(--torch);padding:14px;background:var(--torch-bg);margin-top:10px;}
.loop-box .loop-label{font-family:var(--mono);font-size:8px;letter-spacing:.2em;text-transform:uppercase;color:var(--torch);margin-bottom:8px;}
.loop-steps{display:flex;flex-direction:column;gap:4px;}
.loop-step{display:flex;align-items:center;gap:10px;font-size:12px;padding:5px 8px;background:var(--paper);border-left:2px solid transparent;transition:all .2s;}
.loop-step:hover{border-left-color:var(--torch);background:white;}
.loop-step-num{font-family:var(--mono);font-size:9px;font-weight:700;color:var(--torch);width:18px;flex-shrink:0;}
.loop-step-name{font-weight:600;color:var(--ink);margin-right:4px;}
.loop-step-why{color:var(--ink3);font-style:italic;font-size:11px;}
.gpu-grid{display:grid;grid-template-columns:1fr 48px 1fr;gap:0;background:var(--rule);border:1px solid var(--rule);align-items:center;}
.gpu-side{background:var(--paper);padding:24px;}
.gpu-arrow{background:var(--paper2);display:flex;flex-direction:column;align-items:center;justify-content:center;gap:4px;height:100%;padding:12px 0;}
.gpu-arrow span{font-size:20px;color:var(--torch);animation:pulse-arrow 2s ease-in-out infinite;}
.gpu-arrow small{font-family:var(--mono);font-size:8px;color:var(--ink3);writing-mode:vertical-rl;letter-spacing:.1em;text-transform:uppercase;}
@keyframes pulse-arrow{0%,100%{opacity:.4;}50%{opacity:1;}}
.device-label{font-family:var(--mono);font-size:9px;letter-spacing:.2em;text-transform:uppercase;margin-bottom:12px;padding-bottom:8px;border-bottom:1px solid var(--rule);}
.gpu-side.cpu .device-label{color:var(--blue);}
.gpu-side.gpu .device-label{color:var(--green);}
.device-code{font-family:var(--mono);font-size:10.5px;line-height:1.75;background:var(--paper2);border:1px solid var(--rule);padding:12px 14px;margin-bottom:12px;}
.device-note{font-size:11px;color:var(--ink3);font-style:italic;line-height:1.6;}
.device-tip{margin-top:10px;background:var(--green-bg);border-left:3px solid var(--green);padding:8px 10px;font-size:11px;color:var(--green);font-family:var(--mono);}
.ddp-diagram{border:1px solid var(--rule);background:var(--paper2);padding:24px;}
.ddp-title{font-family:var(--mono);font-size:9px;letter-spacing:.2em;text-transform:uppercase;color:var(--ink3);margin-bottom:20px;}
.ddp-row{display:flex;gap:16px;align-items:flex-start;margin-bottom:16px;}
.ddp-label-col{width:100px;flex-shrink:0;font-family:var(--mono);font-size:9px;color:var(--ink3);text-transform:uppercase;letter-spacing:.08em;padding-top:10px;}
.ddp-gpus{display:flex;gap:8px;flex:1;}
.ddp-gpu{flex:1;border:1.5px solid var(--rule);background:var(--paper);padding:12px 10px;text-align:center;transition:background .2s;}
.ddp-gpu:hover{background:var(--blue-bg);border-color:var(--blue);}
.ddp-gpu-id{font-family:var(--mono);font-size:8px;font-weight:700;color:var(--blue);letter-spacing:.1em;margin-bottom:6px;}
.ddp-gpu-content{font-size:10px;color:var(--ink3);line-height:1.5;}
.ddp-sync-bar{height:28px;background:linear-gradient(90deg,var(--paper) 0%,var(--blue-bg) 50%,var(--paper) 100%);border:1px dashed var(--blue);display:flex;align-items:center;justify-content:center;font-family:var(--mono);font-size:9px;color:var(--blue);letter-spacing:.15em;text-transform:uppercase;margin:4px 0;}
.ddp-note{margin-top:16px;background:var(--paper);border-left:3px solid var(--blue);padding:10px 12px;font-size:11px;color:var(--ink3);font-style:italic;line-height:1.6;}
.ddp-cmd{margin-top:12px;background:var(--ink);color:var(--paper);padding:12px 16px;font-family:var(--mono);font-size:11px;letter-spacing:.04em;}
.ddp-cmd .prompt{color:var(--torch);margin-right:6px;}
.ref-grid{display:grid;grid-template-columns:repeat(2,1fr);gap:2px;background:var(--rule);border:1px solid var(--rule);}
.ref-cell{background:var(--paper);padding:18px 16px;}
.ref-cell-title{font-family:var(--mono);font-size:9px;letter-spacing:.15em;text-transform:uppercase;color:var(--ink3);margin-bottom:10px;padding-bottom:6px;border-bottom:1px solid var(--rule);}
.ref-item{display:flex;gap:8px;align-items:baseline;padding:4px 0;border-bottom:1px solid transparent;transition:border-color .15s;}
.ref-item:hover{border-bottom-color:var(--rule);}
.ref-code{font-family:var(--mono);font-size:10px;color:var(--torch);flex-shrink:0;}
.ref-desc{font-size:11px;color:var(--ink3);}
.footer{margin-top:48px;padding-top:16px;border-top:1px solid var(--rule);display:flex;justify-content:space-between;align-items:center;font-family:var(--mono);font-size:9px;color:var(--ink3);letter-spacing:.08em;}
.note{display:flex;gap:10px;align-items:flex-start;background:var(--gold-bg);border:1px solid #e8d470;padding:12px 14px;margin-top:10px;font-size:12px;line-height:1.6;color:var(--ink2);}
.note-icon{font-size:14px;flex-shrink:0;margin-top:1px;}
@keyframes fadeIn{from{opacity:0;transform:translateY(12px);}to{opacity:1;transform:translateY(0);}}
.three-cols,.tensor-grid,.compgraph-wrap,.pipeline,.gpu-grid,.ddp-diagram,.ref-grid{animation:fadeIn .5s ease both;}
.three-cols{animation-delay:.05s;}.tensor-grid{animation-delay:.10s;}
.compgraph-wrap{animation-delay:.15s;}.pipeline{animation-delay:.20s;}
.gpu-grid{animation-delay:.25s;}.ddp-diagram{animation-delay:.30s;}.ref-grid{animation-delay:.35s;}
</style>
</head>
<body>
<div class="page">
<a href="index.html" class="back-link">← Back to index</a>
<div class="header">
<div class="header-kicker">Visual Reference · Sebastian Raschka, PhD</div>
<h1>PyTorch in <em>One Hour</em></h1>
<div class="header-sub">From tensors to multi-GPU training — every essential concept mapped, annotated, and connected.</div>
<div class="header-meta">
<span class="tag">PyTorch 2.4.1</span>
<span class="tag">9 Sections</span>
<span class="tag">Tensors · Autograd · Training · GPU</span>
<span class="tag">DDP Multi-GPU</span>
</div>
</div>
<!-- sec 1 -->
<div class="sec-header"><span class="sec-num">§ 01</span><span class="sec-title">The Three Core Components of PyTorch</span></div>
<div class="three-cols">
<div class="comp-card tensor">
<div class="comp-num">COMPONENT I</div>
<div class="comp-name">Tensor Library</div>
<div class="comp-sub">torch.Tensor · NumPy-like API</div>
<div class="comp-desc">Multi-dimensional array containers for all data and parameters. GPU-accelerated. The fundamental unit of every PyTorch computation — scalars, vectors, matrices, and beyond.</div>
<div class="comp-analogy">Like NumPy, but your arrays can live on a GPU and know how to differentiate themselves.</div>
</div>
<div class="comp-card autograd">
<div class="comp-num">COMPONENT II</div>
<div class="comp-name">Autograd Engine</div>
<div class="comp-sub">torch.autograd · Dynamic computation graphs</div>
<div class="comp-desc">Automatically computes gradients of any tensor expression. Builds a computation graph on every forward pass and runs backpropagation via <code>.backward()</code> — no calculus by hand.</div>
<div class="comp-analogy">The engine that turns forward passes into backward passes without you writing a single derivative.</div>
</div>
<div class="comp-card dl">
<div class="comp-num">COMPONENT III</div>
<div class="comp-name">Deep Learning Utilities</div>
<div class="comp-sub">torch.nn · Layers, losses, optimizers</div>
<div class="comp-desc">Modular building blocks: <code>nn.Module</code>, <code>nn.Linear</code>, <code>nn.Sequential</code>, loss functions, optimizers (SGD, Adam), data loading (<code>Dataset</code>, <code>DataLoader</code>).</div>
<div class="comp-analogy">A Lego kit for neural networks — mix, match, subclass, and extend.</div>
</div>
</div>
<!-- sec 2 -->
<div class="sec-header"><span class="sec-num">§ 02</span><span class="sec-title">Tensors — The Universal Data Container</span></div>
<div class="tensor-grid">
<div class="tensor-cell">
<div class="tensor-rank">RANK 0 · 0D</div>
<div class="tensor-name">Scalar</div>
<div class="tensor-vis"><div class="scalar-dot"></div></div>
<div class="tensor-code">torch.tensor(42)</div>
<div class="tensor-example">A single loss value, a learning rate</div>
</div>
<div class="tensor-cell">
<div class="tensor-rank">RANK 1 · 1D</div>
<div class="tensor-name">Vector</div>
<div class="tensor-vis"><div class="vec-bar"><span></span><span></span><span></span><span></span></div></div>
<div class="tensor-code">torch.tensor([1, 2, 3, 4])</div>
<div class="tensor-example">A bias term, a 1D feature vector</div>
</div>
<div class="tensor-cell">
<div class="tensor-rank">RANK 2 · 2D</div>
<div class="tensor-name">Matrix</div>
<div class="tensor-vis"><div class="mat-grid"><span></span><span></span><span></span><span></span><span></span><span></span><span></span><span></span><span></span></div></div>
<div class="tensor-code">torch.tensor([[1,2],[3,4]])</div>
<div class="tensor-example">A weight matrix, a batch of embeddings</div>
</div>
<div class="tensor-cell">
<div class="tensor-rank">RANK 3+ · nD</div>
<div class="tensor-name">Tensor</div>
<div class="tensor-vis"><div class="tensor3d-wrap"><div class="tensor3d-layer"></div><div class="tensor3d-layer"></div><div class="tensor3d-layer"></div></div></div>
<div class="tensor-code">shape: [batch, seq, dim]</div>
<div class="tensor-example">A batch of token embeddings for an LLM</div>
</div>
</div>
<table class="dtype-table" style="margin-top:2px;">
<thead><tr><th>dtype</th><th>bits</th><th>created from</th><th>use case</th><th>convert with</th></tr></thead>
<tbody>
<tr><td><span class="dtype-key">torch.float32</span></td><td>32</td><td><span style="color:var(--torch)">Python float</span></td><td>Default for training — GPU-optimized, sufficient precision</td><td><span class="dtype-key">.to(torch.float32)</span></td></tr>
<tr><td><span class="dtype-key">torch.float16 / bfloat16</span></td><td>16</td><td>explicit cast</td><td>Mixed-precision training, LLM inference, saves memory</td><td><span class="dtype-key">.half() / .to(torch.bfloat16)</span></td></tr>
<tr><td><span class="dtype-key">torch.int64</span></td><td>64</td><td><span style="color:var(--torch)">Python int</span></td><td>Class labels, token IDs, indices</td><td><span class="dtype-key">.long()</span></td></tr>
<tr><td><span class="dtype-key">torch.bool</span></td><td>8</td><td>comparison ops</td><td>Attention masks, padding masks</td><td><span class="dtype-key">.bool()</span></td></tr>
</tbody>
</table>
<div class="note" style="margin-top:10px;">
<span class="note-icon">⚠</span>
<span><strong>Key operations:</strong> <code style="font-family:var(--mono);font-size:11px;color:var(--torch)">.shape</code> · <code style="font-family:var(--mono);font-size:11px;color:var(--torch)">.view()/.reshape()</code> · <code style="font-family:var(--mono);font-size:11px;color:var(--torch)">.T</code> (transpose) · <code style="font-family:var(--mono);font-size:11px;color:var(--torch)">@</code> (matmul) · <code style="font-family:var(--mono);font-size:11px;color:var(--torch)">.to(device)</code>. All operations preserve the computation graph when <code style="font-family:var(--mono);font-size:11px;color:var(--torch)">requires_grad=True</code>.</span>
</div>
<!-- sec 3-4 -->
<div class="sec-header"><span class="sec-num">§ 03 – 04</span><span class="sec-title">Computation Graphs & Automatic Differentiation</span></div>
<div class="compgraph-wrap">
<div class="compgraph-panel">
<h3>Forward Pass — building the graph</h3>
<div class="fwd-flow">
<div class="fwd-node">
<div class="fwd-symbol">x₁,w₁</div>
<div class="fwd-label"><strong>Input & Parameters</strong><span>feature tensor + weight with requires_grad=True</span></div>
<div class="fwd-code">tensor([1.1])<br>tensor([2.2],✓grad)</div>
</div>
<div class="fwd-node">
<div class="fwd-symbol">z</div>
<div class="fwd-label"><strong>Net Input</strong><span>linear combination: z = x₁·w₁ + b</span></div>
<div class="fwd-code">z = x1 * w1 + b</div>
</div>
<div class="fwd-node">
<div class="fwd-symbol">a</div>
<div class="fwd-label"><strong>Activation</strong><span>nonlinearity: a = σ(z)</span></div>
<div class="fwd-code">torch.sigmoid(z)</div>
</div>
<div class="fwd-node" style="background:var(--blue-bg);border-color:var(--blue);">
<div class="fwd-symbol" style="color:var(--blue);">L</div>
<div class="fwd-label"><strong>Loss</strong><span>compare prediction to true label y</span></div>
<div class="fwd-code">F.binary_cross_entropy(a,y)</div>
</div>
</div>
</div>
<div class="compgraph-panel" style="background:var(--paper2);">
<h3>Backward Pass — computing gradients</h3>
<div class="bwd-flow">
<div class="bwd-node" style="background:var(--gold-bg);border-color:#e8d470;">
<div class="bwd-symbol">∂L</div>
<div class="bwd-label"><strong>Trigger Backprop</strong><span>PyTorch traverses graph right-to-left</span></div>
<div class="bwd-code">loss.backward()</div>
</div>
<div class="bwd-node">
<div class="bwd-symbol">∂L/∂a</div>
<div class="bwd-label"><strong>Gradient at Activation</strong><span>derivative of BCE w.r.t. sigmoid output</span></div>
<div class="bwd-code">auto-computed</div>
</div>
<div class="bwd-node">
<div class="bwd-symbol">∂L/∂z</div>
<div class="bwd-label"><strong>Gradient at Net Input</strong><span>chain rule through sigmoid</span></div>
<div class="bwd-code">auto-computed</div>
</div>
<div class="bwd-node" style="background:var(--gold-bg);border-color:#e8d470;">
<div class="bwd-symbol">∂L/∂w₁</div>
<div class="bwd-label"><strong>Parameter Gradient</strong><span>how much does the loss change w.r.t. w₁?</span></div>
<div class="bwd-code">w1.grad → −0.0898</div>
</div>
</div>
<div class="chain-rule-box">
<div class="label">Chain Rule</div>
<div class="chain-rule-math">∂L/∂w₁ = (∂L/∂a) · (∂a/∂z) · (∂z/∂w₁)</div>
<div class="chain-rule-note">PyTorch does this automatically. You never compute derivatives by hand.</div>
</div>
</div>
</div>
<!-- sec 5-7 training pipeline -->
<div class="sec-header"><span class="sec-num">§ 05 – 07</span><span class="sec-title">The Full Training Pipeline</span></div>
<div class="pipeline">
<div class="pipe-step">
<div class="pipe-num">1<small>define</small></div>
<div class="pipe-info">
<div class="pipe-name">Dataset</div>
<div class="pipe-desc">Subclass <code>torch.utils.data.Dataset</code>. Implement three methods: <code>__init__</code> (store data), <code>__getitem__</code> (return one example by index), <code>__len__</code> (total size).</div>
<div class="pipe-gotcha">⚠ Class labels must start at 0. Largest label = num_outputs − 1.</div>
</div>
<div class="pipe-code">
<span class="kw">class</span> <span class="fn">MyDataset</span>(Dataset):<br>
<span class="kw">def</span> <span class="fn">__init__</span>(self, X, y):<br>
self.X, self.y = X, y<br>
<span class="kw">def</span> <span class="fn">__getitem__</span>(self, i):<br>
<span class="kw">return</span> self.X[i], self.y[i]<br>
<span class="kw">def</span> <span class="fn">__len__</span>(self):<br>
<span class="kw">return</span> <span class="fn">len</span>(self.y)
</div>
</div>
<div class="pipe-step">
<div class="pipe-num">2<small>load</small></div>
<div class="pipe-info">
<div class="pipe-name">DataLoader</div>
<div class="pipe-desc">Wraps Dataset to handle batching, shuffling, and parallelism. <code>num_workers>0</code> loads next batch in background while GPU trains on current batch.</div>
<div class="pipe-gotcha">⚠ Use <code>drop_last=True</code> to avoid a tiny last batch. Use <code>shuffle=True</code> only for train, not test.</div>
</div>
<div class="pipe-code">
DataLoader(<br>
dataset=train_ds,<br>
batch_size=<span class="num">32</span>,<br>
shuffle=<span class="kw">True</span>,<br>
num_workers=<span class="num">4</span>,<br>
drop_last=<span class="kw">True</span><br>
)
</div>
</div>
<div class="pipe-step">
<div class="pipe-num">3<small>build</small></div>
<div class="pipe-info">
<div class="pipe-name">Model — nn.Module</div>
<div class="pipe-desc">Subclass <code>nn.Module</code>. Define layers in <code>__init__</code>, connect them in <code>forward()</code>. Return raw <strong>logits</strong> — PyTorch loss functions apply softmax/sigmoid internally.</div>
</div>
<div class="pipe-code">
<span class="kw">class</span> <span class="fn">Net</span>(nn.Module):<br>
<span class="kw">def</span> <span class="fn">__init__</span>(self):<br>
<span class="fn">super</span>().__init__()<br>
self.layers = nn.Sequential(<br>
nn.Linear(<span class="num">50</span>,<span class="num">30</span>), nn.ReLU(),<br>
nn.Linear(<span class="num">30</span>,<span class="num">3</span>), <span class="comment"># logits</span><br>
)<br>
<span class="kw">def</span> <span class="fn">forward</span>(self, x):<br>
<span class="kw">return</span> self.layers(x)
</div>
</div>
<div class="pipe-step" style="min-height:200px;">
<div class="pipe-num">4<small>train</small></div>
<div class="pipe-info">
<div class="pipe-name">Training Loop</div>
<div class="pipe-desc">Iterate epochs → batches. Five operations in strict order each batch:</div>
<div class="loop-box">
<div class="loop-label">Per-batch update — repeat for every epoch</div>
<div class="loop-steps">
<div class="loop-step"><span class="loop-step-num">①</span><span class="loop-step-name">Forward pass</span><span class="loop-step-why">— compute logits from features</span></div>
<div class="loop-step"><span class="loop-step-num">②</span><span class="loop-step-name">Compute loss</span><span class="loop-step-why">— F.cross_entropy(logits, labels)</span></div>
<div class="loop-step"><span class="loop-step-num">③</span><span class="loop-step-name">Zero gradients</span><span class="loop-step-why">— optimizer.zero_grad() — prevents accumulation!</span></div>
<div class="loop-step"><span class="loop-step-num">④</span><span class="loop-step-name">Backward pass</span><span class="loop-step-why">— loss.backward() — fills .grad attributes</span></div>
<div class="loop-step"><span class="loop-step-num">⑤</span><span class="loop-step-name">Update params</span><span class="loop-step-why">— optimizer.step() — w ← w − lr·∂L/∂w</span></div>
</div>
</div>
</div>
<div class="pipe-code">
model.<span class="fn">train</span>()<br>
<span class="kw">for</span> features, labels <span class="kw">in</span> loader:<br>
logits = model(features)<br>
loss = F.<span class="fn">cross_entropy</span>(logits, labels)<br>
optimizer.<span class="fn">zero_grad</span>() <span class="comment"># ③</span><br>
loss.<span class="fn">backward</span>() <span class="comment"># ④</span><br>
optimizer.<span class="fn">step</span>() <span class="comment"># ⑤</span>
</div>
</div>
<div class="pipe-step">
<div class="pipe-num">5<small>eval</small></div>
<div class="pipe-info">
<div class="pipe-name">Inference & Save</div>
<div class="pipe-desc"><code>model.eval()</code> disables dropout/batchnorm training behavior. Wrap in <code>torch.no_grad()</code> to skip building the computation graph — saves memory and compute during inference.</div>
<div class="pipe-gotcha">⚠ Always call model.eval() before any prediction. Forgetting this with dropout gives different results each run.</div>
</div>
<div class="pipe-code">
<span class="comment"># Inference</span><br>
model.<span class="fn">eval</span>()<br>
<span class="kw">with</span> torch.<span class="fn">no_grad</span>():<br>
probs = torch.<span class="fn">softmax</span>(model(X), dim=<span class="num">1</span>)<br>
preds = torch.<span class="fn">argmax</span>(probs, dim=<span class="num">1</span>)<br>
<br>
<span class="comment"># Save / Load</span><br>
torch.<span class="fn">save</span>(model.<span class="fn">state_dict</span>(), <span class="str">"model.pth"</span>)<br>
model.<span class="fn">load_state_dict</span>(<br>
torch.<span class="fn">load</span>(<span class="str">"model.pth"</span>, weights_only=<span class="kw">True</span>))
</div>
</div>
</div>
<!-- sec 9.1-9.2 GPU -->
<div class="sec-header"><span class="sec-num">§ 09.1 – 09.2</span><span class="sec-title">GPU Computing — Device Model</span></div>
<div class="gpu-grid">
<div class="gpu-side cpu">
<div class="device-label">CPU (Default)</div>
<div class="device-code">
<span style="color:var(--ink3);"># All tensors start here</span><br>
t = torch.tensor([1., 2., 3.])<br>
<span style="color:var(--ink3);"># t.device → cpu</span>
</div>
<div class="device-note">Every tensor and model parameter starts on CPU. Operations between tensors must be on the same device — mixing CPU and GPU raises <code>RuntimeError</code>.</div>
</div>
<div class="gpu-arrow">
<span>→</span>
<small>.to(device)</small>
</div>
<div class="gpu-side gpu">
<div class="device-label">GPU (cuda / mps)</div>
<div class="device-code">
<span style="color:var(--ink3);"># Best practice device setup</span><br>
device = torch.device(<br>
<span style="color:var(--green);">"cuda"</span> <span style="color:var(--blue);">if</span> torch.cuda.is_available()<br>
<span style="color:var(--blue);">else</span> <span style="color:var(--green);">"cpu"</span><br>
)<br>
model.to(device)<br>
X, y = X.to(device), y.to(device)
</div>
<div class="device-tip">Apple Silicon: replace "cuda" with "mps" — torch.backends.mps.is_available()</div>
</div>
</div>
<div class="note" style="margin-top:10px;">
<span class="note-icon">💡</span>
<span>Only <strong>3 lines change</strong> to go from CPU to single-GPU training: <code style="font-family:var(--mono);font-size:11px;color:var(--torch)">device = ...</code> · <code style="font-family:var(--mono);font-size:11px;color:var(--torch)">model.to(device)</code> · <code style="font-family:var(--mono);font-size:11px;color:var(--torch)">features.to(device), labels.to(device)</code> inside the loop. GPU won’t speed up tiny datasets — transfer overhead dominates.</span>
</div>
<!-- sec 9.3 DDP -->
<div class="sec-header"><span class="sec-num">§ 09.3</span><span class="sec-title">Distributed Training — DistributedDataParallel (DDP)</span></div>
<div class="ddp-diagram">
<div class="ddp-title">How DDP Works — One Process Per GPU, Synchronized Gradients</div>
<div class="ddp-row">
<div class="ddp-label-col">Training Data</div>
<div class="ddp-gpus">
<div class="ddp-gpu"><div class="ddp-gpu-id">BATCH A</div><div class="ddp-gpu-content">examples 1, 3, 5, 7<br><em>DistributedSampler</em><br>ensures no overlap</div></div>
<div class="ddp-gpu"><div class="ddp-gpu-id">BATCH B</div><div class="ddp-gpu-content">examples 2, 4, 6, 8<br><em>different subset</em><br>per GPU process</div></div>
<div class="ddp-gpu" style="border-style:dashed;opacity:.5;"><div class="ddp-gpu-id">BATCH C …</div><div class="ddp-gpu-content">scales linearly<br>with N GPUs</div></div>
</div>
</div>
<div style="display:flex;gap:8px;padding:0 116px;">
<div style="flex:1;text-align:center;font-size:20px;color:var(--ink3);">↓</div>
<div style="flex:1;text-align:center;font-size:20px;color:var(--ink3);">↓</div>
<div style="flex:1;text-align:center;font-size:20px;color:var(--ink3);">↓</div>
</div>
<div class="ddp-row">
<div class="ddp-label-col">Model Copy<br>(identical)</div>
<div class="ddp-gpus">
<div class="ddp-gpu" style="border-color:var(--blue);"><div class="ddp-gpu-id">GPU 0 · rank=0</div><div class="ddp-gpu-content">full model copy<br>forward → loss<br>→ backward → ∇w</div></div>
<div class="ddp-gpu" style="border-color:var(--blue);"><div class="ddp-gpu-id">GPU 1 · rank=1</div><div class="ddp-gpu-content">full model copy<br>forward → loss<br>→ backward → ∇w</div></div>
<div class="ddp-gpu" style="border-color:var(--blue);border-style:dashed;opacity:.5;"><div class="ddp-gpu-id">GPU N …</div><div class="ddp-gpu-content">…</div></div>
</div>
</div>
<div class="ddp-sync-bar">↔ All-Reduce: average gradients across all GPUs (NCCL) ↔</div>
<div class="ddp-row" style="margin-bottom:0;">
<div class="ddp-label-col">Synchronized<br>Weight Update</div>
<div class="ddp-gpus">
<div class="ddp-gpu" style="border-color:var(--green);"><div class="ddp-gpu-id" style="color:var(--green);">GPU 0 · updated</div><div class="ddp-gpu-content">optimizer.step()<br>same weights as GPU 1</div></div>
<div class="ddp-gpu" style="border-color:var(--green);"><div class="ddp-gpu-id" style="color:var(--green);">GPU 1 · updated</div><div class="ddp-gpu-content">optimizer.step()<br>same weights as GPU 0</div></div>
<div class="ddp-gpu" style="border-color:var(--green);border-style:dashed;opacity:.5;"><div class="ddp-gpu-id" style="color:var(--green);">GPU N …</div><div class="ddp-gpu-content">…</div></div>
</div>
</div>
<div class="ddp-note"><strong>The key insight:</strong> each GPU sees a different data subset per iteration, but their gradients are averaged before each weight update — so all model copies stay identical. With 2 GPUs you process 2× more data per wall-clock second. With 8 GPUs, ~8×. Overhead: one all-reduce communication step per iteration.</div>
<div class="ddp-cmd">
<span class="prompt">$</span>torchrun --nproc_per_node=2 train.py <span style="color:var(--ink3);font-size:10px;"># 2 GPUs</span><br>
<span class="prompt">$</span>torchrun --nproc_per_node=$(nvidia-smi -L | wc -l) train.py <span style="color:var(--ink3);font-size:10px;"># all GPUs</span>
</div>
</div>
<!-- REF -->
<div class="sec-header"><span class="sec-num">REF</span><span class="sec-title">Quick Reference Card</span></div>
<div class="ref-grid">
<div class="ref-cell">
<div class="ref-cell-title">Tensor Operations</div>
<div class="ref-item"><span class="ref-code">t.shape</span><span class="ref-desc">dimensions, e.g. torch.Size([2, 3])</span></div>
<div class="ref-item"><span class="ref-code">t.view(3,2)</span><span class="ref-desc">reshape (shares memory)</span></div>
<div class="ref-item"><span class="ref-code">t.T</span><span class="ref-desc">transpose (flip along diagonal)</span></div>
<div class="ref-item"><span class="ref-code">A @ B</span><span class="ref-desc">matrix multiply (= A.matmul(B))</span></div>
<div class="ref-item"><span class="ref-code">t.to(device)</span><span class="ref-desc">move to CPU / GPU / MPS</span></div>
<div class="ref-item"><span class="ref-code">t.item()</span><span class="ref-desc">extract Python scalar from 0D tensor</span></div>
</div>
<div class="ref-cell">
<div class="ref-cell-title">Model & Training</div>
<div class="ref-item"><span class="ref-code">model.train()</span><span class="ref-desc">enable dropout / batchnorm training mode</span></div>
<div class="ref-item"><span class="ref-code">model.eval()</span><span class="ref-desc">disable dropout, freeze batchnorm stats</span></div>
<div class="ref-item"><span class="ref-code">torch.no_grad()</span><span class="ref-desc">skip graph construction (inference)</span></div>
<div class="ref-item"><span class="ref-code">optimizer.zero_grad()</span><span class="ref-desc">clear gradients before each backward</span></div>
<div class="ref-item"><span class="ref-code">loss.backward()</span><span class="ref-desc">compute all gradients via chain rule</span></div>
<div class="ref-item"><span class="ref-code">optimizer.step()</span><span class="ref-desc">update parameters using computed gradients</span></div>
</div>
<div class="ref-cell">
<div class="ref-cell-title">Model Persistence</div>
<div class="ref-item"><span class="ref-code">model.state_dict()</span><span class="ref-desc">dict of all parameter tensors</span></div>
<div class="ref-item"><span class="ref-code">torch.save(sd, "f.pth")</span><span class="ref-desc">save to disk</span></div>
<div class="ref-item"><span class="ref-code">torch.load("f.pth", weights_only=True)</span><span class="ref-desc">load dict</span></div>
<div class="ref-item"><span class="ref-code">model.load_state_dict(sd)</span><span class="ref-desc">restore weights (architecture must match)</span></div>
<div class="ref-item"><span class="ref-code">sum(p.numel() for p in model.parameters())</span><span class="ref-desc">total params</span></div>
</div>
<div class="ref-cell">
<div class="ref-cell-title">Common Gotchas</div>
<div class="ref-item"><span class="ref-code">zero_grad()</span><span class="ref-desc">must call before backward() — gradients accumulate by default</span></div>
<div class="ref-item"><span class="ref-code">logits (not softmax)</span><span class="ref-desc">return raw logits from model — loss fns apply softmax internally</span></div>
<div class="ref-item"><span class="ref-code">same device</span><span class="ref-desc">model and data must be on identical device or RuntimeError</span></div>
<div class="ref-item"><span class="ref-code">drop_last=True</span><span class="ref-desc">prevents tiny last batch destabilizing gradient updates</span></div>
<div class="ref-item"><span class="ref-code">shuffle=False</span><span class="ref-desc">test/val loaders must not shuffle — use DistributedSampler for DDP</span></div>
</div>
</div>
<div class="footer">
<span>SOURCE: SEBASTIANRASCHKA.COM/TEACHING/PYTORCH-1H · JUL 1, 2025</span>
<span>VIZ: CLAUDE · FEB 2026</span>
</div>
</div>
</body>
</html>