File size: 9,779 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 |
WEBVTT X-TIMESTAMP-MAP=LOCAL:00:00:00.000,MPEGTS:144533 1 00:00:02.002 --> 00:00:03.003 Now, let's go 2 00:00:03.003 --> 00:00:05.972 even smaller and do per group quantization. 3 00:00:05.972 --> 00:00:10.777 In per group quantization we perform quantization on groups of n elements. 4 00:00:11.144 --> 00:00:13.813 Common values for n are 32, 5 00:00:13.813 --> 00:00:16.883 64, or 128. Per group 6 00:00:16.883 --> 00:00:19.619 quantization can require a lot of memory. 7 00:00:19.619 --> 00:00:20.086 Let's say, 8 00:00:20.086 --> 00:00:25.086 we want to quantize a tensor in four-bit, and we choose a group size equal to 32. 9 00:00:25.325 --> 00:00:26.459 We use symmetric mode. 10 00:00:26.459 --> 00:00:29.062 That means that the zero point is equal to zero, 11 00:00:29.062 --> 00:00:32.265 and we store the scales in floating point 16. 12 00:00:32.599 --> 00:00:37.504 It means that we are actually quantizing the tensor in 4.5 bits. 13 00:00:37.871 --> 00:00:42.709 Since we have four bits, since each element is stored using four bit 14 00:00:43.343 --> 00:00:47.414 and we have 16 divided by 32 bit. 15 00:00:47.747 --> 00:00:51.985 Since we need to store a scale in 16 bits 16 00:00:51.985 --> 00:00:56.823 for every 32 elements for each element, you store it in four bit, 17 00:00:56.923 --> 00:01:01.923 but you also have quantization parameters and you need to store once 18 00:01:02.462 --> 00:01:07.033 a scale in 16 bits, so 16 bits every 32 elements. 19 00:01:07.167 --> 00:01:09.069 Now let's jump to the code. 20 00:01:09.069 --> 00:01:12.305 For simplicity, we will restrict ourselves to the case 21 00:01:12.305 --> 00:01:16.709 where the tensor is of dimension two and we will be using the symmetric mode. 22 00:01:16.943 --> 00:01:17.210 You don't 23 00:01:17.210 --> 00:01:20.680 need to pay attention to this code since we will be coding in the notebook. 24 00:01:20.980 --> 00:01:22.615 Now let's code it. 25 00:01:22.615 --> 00:01:25.518 So we define the following function. 26 00:01:25.518 --> 00:01:27.353 Linear q symmetric per group. 27 00:01:28.955 --> 00:01:31.958 This will take as argument the tensor, 28 00:01:33.126 --> 00:01:36.129 the group size and the d-type. 29 00:01:37.797 --> 00:01:41.000 We set the default value torch.int8. 30 00:01:42.502 --> 00:01:46.239 First, we need to get the shape of the tensor. 31 00:01:50.276 --> 00:01:53.413 Then, another restriction for this function 32 00:01:53.613 --> 00:01:57.750 is that we will be performing quantization on the rows. 33 00:01:57.851 --> 00:02:02.851 This is why we also need to make sure that each row is divisible by group size. 34 00:02:02.922 --> 00:02:05.892 To confirm that, we will just use assertion 35 00:02:05.892 --> 00:02:10.892 so that the shape of the tensor along the rows is indeed divisible by group size. 36 00:02:13.366 --> 00:02:17.637 Then, as I said, we will be restricting ourselves 37 00:02:17.637 --> 00:02:20.640 to tensors of dimension two. 38 00:02:22.509 --> 00:02:26.646 Now all we need to do is to reshape the tensor 39 00:02:26.980 --> 00:02:31.084 so that we end up with rows of group size elements. 40 00:02:31.784 --> 00:02:36.122 To do that, we will use the view function that we learned about. 41 00:02:40.527 --> 00:02:42.862 So as you can see, what we do here 42 00:02:42.862 --> 00:02:46.499 is to make sure that each row contains group size elements. 43 00:02:46.799 --> 00:02:49.702 And we put the minus one here so that it infers 44 00:02:49.702 --> 00:02:53.273 automatically the right dimension to have in the first dimension. 45 00:02:53.540 --> 00:02:57.310 And now if you look at the tensor we has the setup 46 00:02:57.577 --> 00:03:00.413 for performing positional quantization. 47 00:03:00.413 --> 00:03:04.984 We resized this tensor so that we have rows of group size 48 00:03:05.084 --> 00:03:09.088 so that we can use the function that we coded previously. 49 00:03:09.522 --> 00:03:12.959 That is to say the linear q symmetric channel quantization. 50 00:03:13.126 --> 00:03:16.129 So we have quantized tensor 51 00:03:16.763 --> 00:03:17.864 and scale 52 00:03:17.864 --> 00:03:21.768 which is equal to linear q symmetric per channel function. 53 00:03:22.669 --> 00:03:26.506 And we need to put the tensor the right dimension. 54 00:03:26.906 --> 00:03:29.976 So along the rows and the d-type. 55 00:03:33.179 --> 00:03:35.348 After quantizing the tensor 56 00:03:35.348 --> 00:03:38.551 we still need to reshape it to its original shape. 57 00:03:38.851 --> 00:03:41.688 So we will use the shape that we stored before. 58 00:03:41.688 --> 00:03:43.022 Here the d shape. 59 00:03:46.292 --> 00:03:47.594 To reshape the tensor, 60 00:03:47.594 --> 00:03:51.130 we use the view and we just pass this shape. 61 00:03:51.564 --> 00:03:54.567 Then we can return the quantized tensor and the scale. 62 00:03:55.902 --> 00:03:57.570 Now that we have coded 63 00:03:57.570 --> 00:04:02.342 the per group quantization, now let's code the linear quantization 64 00:04:02.575 --> 00:04:07.380 for the quantization in order to verify our results. 65 00:04:07.947 --> 00:04:10.950 So we need to define 66 00:04:10.950 --> 00:04:11.951 this function. 67 00:04:11.951 --> 00:04:15.021 In that function we need the quantized tensor 68 00:04:16.022 --> 00:04:18.658 to scale. 69 00:04:18.658 --> 00:04:21.661 But we also need the group size. 70 00:04:25.098 --> 00:04:29.068 Then we need to get the shape of the quantized tensor. 71 00:04:29.102 --> 00:04:32.105 That will be useful. 72 00:04:33.106 --> 00:04:35.842 Then we need to reshape 73 00:04:35.842 --> 00:04:39.479 the quantized tensor so that we have rows that contain 74 00:04:39.479 --> 00:04:42.482 only group size elements. 75 00:04:42.548 --> 00:04:46.986 To do that, we put in the view methods minus 76 00:04:46.986 --> 00:04:51.190 one for the first value and group size for the second one. 77 00:04:52.191 --> 00:04:56.062 Then we can reuse the linear 78 00:04:56.162 --> 00:05:00.366 dequantization methods we coded before to dequantize the tensor. 79 00:05:00.600 --> 00:05:05.600 We need to pass the quantized tensor, the scale and decimal point. 80 00:05:06.072 --> 00:05:08.641 But since we are doing symmetric quantization, 81 00:05:09.942 --> 00:05:11.878 the zero point is equal to zero. 82 00:05:11.878 --> 00:05:16.382 Then all we need to do is to reshape the dequantized tensor 83 00:05:16.749 --> 00:05:21.554 with the shape of the original tensor, and the shape is stored in q shape. 84 00:05:24.223 --> 00:05:27.226 Then we return the dequantized tensor. 85 00:05:27.994 --> 00:05:30.997 Now let's test our implementation. 86 00:05:30.997 --> 00:05:33.933 We will test a random tensor of size six 87 00:05:33.933 --> 00:05:36.903 by six and 88 00:05:37.937 --> 00:05:40.840 let's set group size to be equal to three. 89 00:05:40.840 --> 00:05:45.478 So, we will get the quantized tensor and the scale 90 00:05:46.546 --> 00:05:50.450 using the linear q symmetric group function. 91 00:05:50.850 --> 00:05:53.553 And we need to pass the test 92 00:05:53.553 --> 00:05:56.556 tensor as well as the group size. 93 00:05:56.756 --> 00:06:00.293 Then to verify our results we also need to 94 00:06:00.293 --> 00:06:03.296 dequantize the tensor using 95 00:06:03.296 --> 00:06:07.500 the linear dequantization function where we need to pass 96 00:06:08.101 --> 00:06:10.737 the quantized tensor, 97 00:06:10.737 --> 00:06:13.306 the scale, and the group size. 98 00:06:13.306 --> 00:06:16.743 Finally, to have the summary of the quantization process, 99 00:06:17.377 --> 00:06:20.380 we just need to pass inside the plot quantization error. 100 00:06:20.813 --> 00:06:23.216 The following arguments. 101 00:06:23.216 --> 00:06:26.419 So test tensor, quantized tensor and dequantized tensor. 102 00:06:26.853 --> 00:06:31.224 And as you can see, if you look at the quantized tensor 103 00:06:31.557 --> 00:06:34.761 you will see that every three elements in each row 104 00:06:35.161 --> 00:06:38.364 you will have the maximum value 127. 105 00:06:38.965 --> 00:06:42.135 It shows that we indeed managed to quantize 106 00:06:42.435 --> 00:06:45.738 each three elements in this matrix along the rows. 107 00:06:45.905 --> 00:06:49.442 So three elements here, three here, and so on. 108 00:06:49.742 --> 00:06:51.411 And you have the quantized tensor. 109 00:06:51.411 --> 00:06:52.979 As you can see on the right. 110 00:06:52.979 --> 00:06:57.979 And you can see also that the quantization error tensor is very, very low 111 00:06:58.518 --> 00:07:01.521 and that the dequantized tensor is 112 00:07:01.988 --> 00:07:04.991 practically the same as the original tensor. 113 00:07:05.057 --> 00:07:09.262 Let's also print the dequantization error using the dequantization error function. 114 00:07:09.529 --> 00:07:12.632 And we just need to pass the test tensor and the quantized tensor. 115 00:07:13.933 --> 00:07:16.903 And indeed we have a very very low quantization error. 116 00:07:16.903 --> 00:07:20.573 Now is a good time to pause the video and try a couple of things. 117 00:07:20.606 --> 00:07:23.042 You can try to change the test tensors. 118 00:07:23.042 --> 00:07:25.311 Or you can also change the group size. 119 00:07:25.311 --> 00:07:30.311 And to see what is the effect of the group size on the dequantization process. |