File size: 5,110 Bytes
0c08f5a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
WEBVTT
X-TIMESTAMP-MAP=LOCAL:00:00:00.000,MPEGTS:144533

1
00:00:00.000 --> 00:00:03.100
Quantization methods are used to make models smaller,

2
00:00:03.100 --> 00:00:06.200
which makes them more accessible to the AI community.

3
00:00:06.200 --> 00:00:09.800
In this lesson, you'll get an overview of what Quantization is,

4
00:00:09.800 --> 00:00:12.500
and how it works. Let's get started.

5
00:00:12.500 --> 00:00:15.500
We have seen previously
that quantization is an exciting topic

6
00:00:15.500 --> 00:00:19.733
as it enables us to shrink models
for better accessibility to the community.

7
00:00:19.900 --> 00:00:22.000
In this lesson,
we will learn how to implement

8
00:00:22.000 --> 00:00:25.400
some quantization primitives from scratch,
and we will also implement

9
00:00:25.400 --> 00:00:29.933
our own model quantizer and cover
some challenges that anyone can face

10
00:00:29.933 --> 00:00:33.700
when it comes to lower bit quantization,
such as weights packing.

11
00:00:34.300 --> 00:00:35.433
Let's get started.

12
00:00:35.433 --> 00:00:38.300
So let's first have a quick glance
on what we have learned

13
00:00:38.300 --> 00:00:40.200
so far from the first course.

14
00:00:40.200 --> 00:00:44.100
So, in the introduction of the first
course, we listed all available techniques

15
00:00:44.100 --> 00:00:47.433
that one could use in order
to compress a model in general.

16
00:00:48.000 --> 00:00:50.733
So first of all, quantization
aims at representing

17
00:00:50.733 --> 00:00:53.733
parameters of the model
in a lower precision.

18
00:00:53.900 --> 00:00:57.700
With knowledge distillation,
you can train a student model

19
00:00:57.700 --> 00:01:00.700
using the bigger teacher model outputs.

20
00:01:00.833 --> 00:01:04.300
And finally, with pruning,
you can simply remove some connections

21
00:01:04.300 --> 00:01:09.300
inside the model, meaning removing weights
to make the model more sparse.

22
00:01:10.100 --> 00:01:12.533
We also covered
common data types in machine

23
00:01:12.533 --> 00:01:16.700
learning, such as INT8 or float.
We also performed

24
00:01:16.800 --> 00:01:21.100
linear quantization using Hugging Face's
quantum library with few lines of code.

25
00:01:21.400 --> 00:01:24.200
And finally,
we wrapped up the course with an overview

26
00:01:24.200 --> 00:01:27.200
how quantization can be leveraged
in different use cases,

27
00:01:27.233 --> 00:01:29.900
such as large language
models, finetuning.

28
00:01:29.900 --> 00:01:32.900
So let's see together what we are going
to cover exactly in this course.

29
00:01:33.333 --> 00:01:37.433
So, first of all, we are going
to deep dive together into the internals

30
00:01:37.433 --> 00:01:41.733
of linear quantization and implement
some of their variants from scratch,

31
00:01:42.000 --> 00:01:46.000
such as per channel,
per tensor or per group quantization.

32
00:01:46.700 --> 00:01:50.200
We will study what are the advantages
and drawbacks for each of these methods,

33
00:01:50.200 --> 00:01:53.100
and we will see their impact
on some random tensors.

34
00:01:53.100 --> 00:01:56.700
And next,
we will try to build our own quantizer

35
00:01:56.900 --> 00:01:58.800
to quantize any model in eight-bit

36
00:01:58.800 --> 00:02:02.500
precision using one of the quantization
schemes presented before.

37
00:02:02.833 --> 00:02:06.200
Note the quantization scheme
is agnostic to modalities,

38
00:02:06.433 --> 00:02:11.000
meaning you can apply to any model as long
as your model contains linear layers.

39
00:02:11.400 --> 00:02:14.400
Technically, you will be able
to use your quantizer to quantize

40
00:02:14.400 --> 00:02:17.900
a vision, text, audio,
or even a multimodal model.

41
00:02:18.200 --> 00:02:21.433
And finally, we will wrap up the course
by learning more

42
00:02:21.500 --> 00:02:26.033
about some challenges that you can face
when it comes to extreme quantization,

43
00:02:26.033 --> 00:02:29.533
such as weight packing,
which is a common challenge these days.

44
00:02:30.000 --> 00:02:32.933
As of the time we speak, PyTorch
does not have

45
00:02:32.933 --> 00:02:36.100
a native support for two-bit or four-bit
precision weights.

46
00:02:36.400 --> 00:02:40.000
One way to address
this issue is to pack these low precision

47
00:02:40.000 --> 00:02:43.433
weights into a higher precision tensor,
for example INT8.

48
00:02:43.833 --> 00:02:45.600
And we will deep dive into that

49
00:02:45.600 --> 00:02:49.300
and implement
packing and unpacking algorithms together.

50
00:02:49.400 --> 00:02:53.300
And we will end the course by covering
what are the other common challenges

51
00:02:53.300 --> 00:02:56.333
when it comes to quantizing large models
such as LLMs.

52
00:02:56.500 --> 00:03:00.600
And review together some state of the art
quantization methods together.

53
00:03:00.833 --> 00:03:03.600
Yeah, so let's try to get started
and shrink some models.