aimlbd / HuggingFace.Quantization_in_Depth /Transcript /14_Packing 2-bit Weights_HFQD.txt
swsuws's picture
Upload 200 files
0c08f5a verified
raw
history blame
12.4 kB
WEBVTT
X-TIMESTAMP-MAP=LOCAL:00:00:00.000,MPEGTS:144533
1
00:00:02.068 --> 00:00:02.302
Okay.
2
00:00:02.302 --> 00:00:04.337
So we're going to do that in pure PyTorch.
3
00:00:04.337 --> 00:00:08.842
Let's first import torch and then define
the signature of the method pack_weights.
4
00:00:09.275 --> 00:00:14.275
That is going to take the unsigned
int8 tensor and the number of bits.
5
00:00:14.514 --> 00:00:19.514
So the reason why we're using unsigned int
instead of int8 is that for int8 tensor,
6
00:00:20.120 --> 00:00:24.090
the first bit is used to determine
the sign of the tensor.
7
00:00:24.224 --> 00:00:27.427
So, just for simplicity
we're going to use unsigned int
8
00:00:27.961 --> 00:00:30.930
so that we don't have to deal
with the sign bit.
9
00:00:30.964 --> 00:00:33.600
So recall also
what I've said at the beginning.
10
00:00:33.600 --> 00:00:38.600
So for simplicity it's better
to have, the shape of the input tensor
11
00:00:39.339 --> 00:00:42.942
being a multiple of four eight
divided by number of bits.
12
00:00:43.209 --> 00:00:46.046
So we're just going to add
a small condition here checking
13
00:00:46.046 --> 00:00:49.049
that if the tensor does
not have the right shape,
14
00:00:49.049 --> 00:00:53.720
then we're going to raise a small error
saying that the input shape needs
15
00:00:53.720 --> 00:00:58.058
to be a multiple of the expected number
and print it to the users.
16
00:00:58.625 --> 00:00:59.759
So, now let's get the
17
00:01:00.794 --> 00:01:01.294
number of
18
00:01:01.294 --> 00:01:04.297
values
expected values of the packed tensor.
19
00:01:04.564 --> 00:01:06.933
So if we come back to our example.
20
00:01:06.933 --> 00:01:10.870
So this of course
depends on the number of bits that you use
21
00:01:10.870 --> 00:01:12.405
to encode your values.
22
00:01:12.405 --> 00:01:15.542
So in case of two bits
we have four values.
23
00:01:15.742 --> 00:01:19.712
So the number of expected packed values
is simply going to be
24
00:01:20.246 --> 00:01:24.651
the shape of your input tensor
multiplied by the number of bits.
25
00:01:24.751 --> 00:01:27.520
So here we have two bits per value.
26
00:01:27.520 --> 00:01:28.621
And then we have four values.
27
00:01:28.621 --> 00:01:32.158
So eight bit divided by eight
28
00:01:33.493 --> 00:01:36.496
because our packed tensor
29
00:01:36.796 --> 00:01:37.997
is going to be eight bits.
30
00:01:37.997 --> 00:01:42.001
So number of expected value
is simply going to be number of input
31
00:01:42.302 --> 00:01:46.106
low bits values
that you're going to retrieve with dot
32
00:01:46.139 --> 00:01:49.142
shape of zero times the number of bits.
33
00:01:49.142 --> 00:01:51.010
If either you are in 2 or 4 bits,
34
00:01:52.245 --> 00:01:55.248
everything divided by eight.
35
00:01:56.716 --> 00:01:57.550
Perfect.
36
00:01:57.550 --> 00:02:00.553
So let's come back again
to the example that we had before.
37
00:02:00.587 --> 00:02:05.587
So in case of two bits, each packet weight
will contain four values two bit.
38
00:02:06.326 --> 00:02:09.729
We're processing
four parameters per packed parameter.
39
00:02:10.096 --> 00:02:13.733
So this is going to be our number
of processing steps.
40
00:02:14.100 --> 00:02:17.370
So we're calling this value num steps.
41
00:02:17.403 --> 00:02:20.373
That is simply going to be
the number of bits
42
00:02:20.373 --> 00:02:23.243
of the pack tensor
divided by number of bits.
43
00:02:23.243 --> 00:02:25.979
Again in our example this should be
44
00:02:25.979 --> 00:02:30.917
four because of a processing
for two bit values for packed tensor.
45
00:02:31.050 --> 00:02:33.887
And we're going to
46
00:02:33.887 --> 00:02:36.723
declare this index
47
00:02:36.723 --> 00:02:38.725
because we're going to have a four loop.
48
00:02:38.725 --> 00:02:43.725
And then we initialize our packed tensor
which should contain num values
49
00:02:44.531 --> 00:02:47.767
and which should be in a
dtype of unsigned eight.
50
00:02:48.168 --> 00:02:48.868
Perfect.
51
00:02:48.868 --> 00:02:51.871
Let's first loop on each packed value.
52
00:02:52.172 --> 00:02:56.543
And then for each packed
value we pack num steps,
53
00:02:57.810 --> 00:02:59.345
low bit values.
54
00:02:59.345 --> 00:03:02.982
So we need to loop again here
using a new variable.
55
00:03:03.516 --> 00:03:08.154
And then we're going to use unpacked index
in order to keep track of
56
00:03:08.688 --> 00:03:11.925
which values
are we trying to pack in our algorithm.
57
00:03:12.525 --> 00:03:15.728
So, in the first iteration again
58
00:03:15.728 --> 00:03:18.731
if we consider the tensor
that we had in our example.
59
00:03:18.932 --> 00:03:20.934
So I'm going to write it here.
60
00:03:20.934 --> 00:03:23.937
So 1,0,3,2
61
00:03:24.137 --> 00:03:26.673
which should be in two bit.
62
00:03:26.673 --> 00:03:29.676
And then we have our
63
00:03:29.809 --> 00:03:31.878
so this is uint8 tensor.
64
00:03:31.878 --> 00:03:33.479
So it's encoded in Int8.
65
00:03:33.479 --> 00:03:36.349
And we only want to extract those bits.
66
00:03:36.349 --> 00:03:40.620
And then we have our plain packed tensor
which should only contain
67
00:03:41.254 --> 00:03:43.356
one value in uint8.
68
00:03:43.356 --> 00:03:44.057
All right.
69
00:03:44.057 --> 00:03:46.593
For each num steps here.
70
00:03:46.593 --> 00:03:48.661
So for each,
71
00:03:48.661 --> 00:03:50.463
num step.
72
00:03:50.463 --> 00:03:54.300
So each two bits, we're going to retrieve
the corresponding value.
73
00:03:54.667 --> 00:03:56.402
So here for example it should be one.
74
00:03:58.404 --> 00:03:58.805
And then,
75
00:03:58.805 --> 00:04:01.941
we're going to perform bitwise
shifting on the left
76
00:04:02.642 --> 00:04:05.645
for this tensor but encoded in eight bits.
77
00:04:06.012 --> 00:04:08.514
So let me try to break it down below.
78
00:04:08.514 --> 00:04:10.416
So here,
79
00:04:10.416 --> 00:04:12.752
this value.
80
00:04:12.752 --> 00:04:14.387
So it's encoded in uint8.
81
00:04:14.387 --> 00:04:17.290
So it should give us this value.
82
00:04:17.290 --> 00:04:18.424
All right.
83
00:04:18.424 --> 00:04:21.661
And then the idea is
that we're going to take this value
84
00:04:22.195 --> 00:04:26.366
shifted on the left by bits times J.
85
00:04:26.766 --> 00:04:30.970
So here since we are on
the first iteration it's going to be zero.
86
00:04:31.170 --> 00:04:33.106
So nothing is going to be applied here.
87
00:04:33.106 --> 00:04:35.375
So no shifting on the left.
88
00:04:35.375 --> 00:04:39.245
So this value should stay like this.
89
00:04:39.979 --> 00:04:40.913
All right.
90
00:04:40.913 --> 00:04:43.916
And then we're going to perform bitwise
91
00:04:44.550 --> 00:04:49.550
or operation on the current packed tensor.
92
00:04:49.722 --> 00:04:50.990
So let me explain you that.
93
00:04:50.990 --> 00:04:55.028
So in the first iteration this is how
the packed tensor would look like.
94
00:04:55.595 --> 00:04:59.499
And this is how the right side
of the equation would look like.
95
00:04:59.499 --> 00:04:59.932
Here.
96
00:05:00.900 --> 00:05:03.903
And again we want to pack this here.
97
00:05:04.037 --> 00:05:07.340
Then this here, this here, and this here.
98
00:05:07.807 --> 00:05:10.843
So we're going to perform
an bitwise or operation.
99
00:05:11.177 --> 00:05:14.681
So that zero and 0 or 1 would give us one
100
00:05:14.681 --> 00:05:17.684
0 or 0 zero and so on.
101
00:05:17.850 --> 00:05:20.420
So here after the first iteration
the packed tensor
102
00:05:20.420 --> 00:05:23.423
would exactly look like this. Okay.
103
00:05:23.423 --> 00:05:26.726
And then on the second iteration
then we're going to increment
104
00:05:26.993 --> 00:05:29.996
our unpacked index here.
105
00:05:31.297 --> 00:05:32.265
All right.
106
00:05:32.265 --> 00:05:36.002
And then on the second iteration
we're going to take this tensor
107
00:05:36.002 --> 00:05:39.005
but encoded in uint8.
108
00:05:41.374 --> 00:05:42.942
So now the packed tensor
109
00:05:42.942 --> 00:05:46.179
would look like this
because of the bitwise or operation.
110
00:05:46.713 --> 00:05:51.017
And this time the shifting coefficient
is going to be two.
111
00:05:51.784 --> 00:05:53.753
So you're going to take this tensor.
112
00:05:53.753 --> 00:05:55.088
It's 000.
113
00:05:55.088 --> 00:05:57.357
Shift it on the left by two.
114
00:05:57.357 --> 00:05:59.792
So again it's going to be 000.
115
00:05:59.792 --> 00:06:02.995
And then you're going to perform
bitwise Or operation
116
00:06:02.995 --> 00:06:06.899
between the shifted tensor
and the pack tensor.
117
00:06:07.300 --> 00:06:10.703
And it's still going to be 1000000.
118
00:06:11.070 --> 00:06:14.941
That's fine because we pack
the first tensor here in two bit.
119
00:06:15.241 --> 00:06:17.143
We pack the second
thing off here in two bits,
120
00:06:18.344 --> 00:06:20.279
and then
we're ready to move on to the next one.
121
00:06:20.279 --> 00:06:23.282
And then on the next iteration
122
00:06:23.383 --> 00:06:25.685
we're going to have this tensor.
123
00:06:25.685 --> 00:06:28.287
Okay. Again so encode you in uint8.
124
00:06:28.287 --> 00:06:30.790
We're going to shift it on the left
this time by.
125
00:06:30.790 --> 00:06:32.725
So J is going to be equal to two.
126
00:06:32.725 --> 00:06:35.061
So two times two four.
127
00:06:35.061 --> 00:06:37.330
So we're going to shift that by four bit.
128
00:06:37.330 --> 00:06:39.499
And it's going to look like this.
129
00:06:39.499 --> 00:06:42.602
And then you're going to perform
bitwise or operation.
130
00:06:42.935 --> 00:06:45.872
And the new pack tensor would look like
131
00:06:45.872 --> 00:06:48.341
this. Perfect.
132
00:06:48.341 --> 00:06:50.676
And then on the last iteration
you'll do the same thing
133
00:06:50.676 --> 00:06:54.180
but this time shifting by six on the left.
134
00:06:54.747 --> 00:06:57.183
So one zero would be here.
135
00:06:57.183 --> 00:07:01.921
And then bitwise or at the very end
you would end up like this.
136
00:07:02.422 --> 00:07:07.059
So the final packed tensor
would look theoretically like this.
137
00:07:11.664 --> 00:07:12.632
Perfect.
138
00:07:12.632 --> 00:07:12.932
Yeah.
139
00:07:12.932 --> 00:07:16.335
Let's try out quickly
our methods and test it.
140
00:07:16.669 --> 00:07:19.472
And we're going to test it on our toy
example.
141
00:07:19.472 --> 00:07:23.276
Let's say our unpacked tensor is encoded
as follows:
142
00:07:23.776 --> 00:07:26.345
1032.
143
00:07:26.345 --> 00:07:28.981
Again everything encoded in two bit.
144
00:07:28.981 --> 00:07:32.218
And yeah let's pack the weights
to see if it works.
145
00:07:32.452 --> 00:07:33.152
Perfect.
146
00:07:33.152 --> 00:07:36.222
So 177 should be encoded
147
00:07:36.856 --> 00:07:39.459
exactly like this in uint8.
148
00:07:39.459 --> 00:07:41.661
You can try that out
verifying the results.
149
00:07:41.661 --> 00:07:43.930
But yeah that should be
the correct result.
150
00:07:43.930 --> 00:07:46.899
Perfect. So yeah.
151
00:07:46.899 --> 00:07:48.701
You can pause the video
152
00:07:48.701 --> 00:07:51.571
and maybe try to understand
this whole logic.
153
00:07:51.571 --> 00:07:53.406
Try also maybe to enhance it a bit.
154
00:07:53.406 --> 00:07:54.841
Optimize it.
155
00:07:54.841 --> 00:07:58.611
You can also try out with four bits
and yeah, maybe you can also try out
156
00:07:58.611 --> 00:08:00.213
different combinations.
157
00:08:00.213 --> 00:08:03.216
For example, you can also battle test
158
00:08:03.249 --> 00:08:07.053
the method a bit if we add these values.
159
00:08:07.353 --> 00:08:09.388
So three and two which should be one one.
160
00:08:09.388 --> 00:08:12.391
So we should have 11111 here everywhere.
161
00:08:12.625 --> 00:08:15.561
And the second tensor should be 255
162
00:08:16.562 --> 00:08:17.363
which is the case.