File size: 25,910 Bytes
370453e |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 |
# GPT2 Experiments
Scripts and logs of GPT2 experiments on Jean Zay HPC.
Using 4x VT100 32GB nodes.
(add `-C v100-32g` for 32gb nodes.)
## Apples and Oranges
JZ seems to give us inconsistent performance - so each allocation may give performance that can vary as much as 40%, so the numbers in the summaries of this document are very hard to compare. We thought it had to do with the proximity of the allocated nodes but it proved to vary randomly through the day, most likely highly dependening on the traffic on the JZ network.
Therefore any results you will find in this summary are +/-40% correct. An identical test scored 40% faster or slower on the same allocation at different times of the day.
## Megatron-LM
Constants:
- `TP_SIZE` = tensor parallel
- `PP_SIZE` = pipeline parallel
- `DP_SIZE` = data parallel is derived automatically from `WORLD_SIZE / (TP_SIZE * PP_SIZE)`
- `WORLD_SIZE` = total number of GPUs
According to Megatron-LM paper the highest degree of TP we can use is 4 for 4-gpu nodes - crossing nodes would slow things down a lot. So max `TP_SIZE=4`. So the full 4 gpu node is used only for tensor parallel dimension.
## Metrics
TFlops: `model_size_in_B * 4 * 2 * seq * global_batch_size / (time_in_sec_per_interation * total_gpus * 1e3)`
The factor of 4 is when used with activation check-pointing,
otherwise it will be 3, but for 200B model, activation check-pointing will always be on.
The peak of V100 32gb gpu is about 125 TFlops/sec [spec](https://images.nvidia.com/content/technologies/volta/pdf/volta-v100-datasheet-update-us-1165301-r5.pdf). But we cannot get the peak. The max achievable performance will be 30-60TFlops depending on the model size. So if you see low 20s, the model is not tuned well, if you see, over 100 then there is a bug in the calculation. 
For v100 16gb gpus the max spec is 120 TFlops/sec.
## Allocation
```
salloc --constraint=v100-32g --nodes=16 --ntasks=16 --cpus-per-task=40 --gres=gpu:4 --hint=nomultithread --time=6:00:00 bash --rcfile $six_ALL_CCFRWORK/start-prod
```
### Megatron
The full slurm scripts and log files are at [`gpt2-meg`](./gpt2-meg):
- scripts starting with `meg_gpt2_base_` are for getting the baseline with tiny BS
- scripts starting with `meg_gpt2_perf_` are for smaller model, and tuned for high performance
Not yet optimized with NVIDIA team!
Metrics can be calculated in bash after figuring out the throughput (in seconds):
```
THROUGHPUT=122
NNODES=16
MSIZE=52
MICRO_BATCH_SIZE=4
DP_SIZE=1
PP_CHUNKS=256
echo "($MSIZE*4*2*1024*$MICRO_BATCH_SIZE*$DP_SIZE*$PP_CHUNKS)/($THROUGHPUT*$NNODES*4*1000)" | bc -l
55.86675409836065573770
```
**Max model size**
These first results are all about how big of a model can be fit into the given the hardware on the smallest batch size, disregarding throughput.
16GB nodes:
| GPUs | Size | DP | PP | PP Chunks | Mic-BS | Glob-BS | Speed | TFlops |
| ---: | ---: | -: | -: | --------: | -----: | -----: | -----: | -----: |
| 16 | 7.5B | 1 | 4 | 4 | 1 | 4 | 0.661s | 23.2 |
| 64 | 30B | 1 | 16 | 4 | 1 | 4 | 1.439s | 10.7 |
| 128 | 50B | 1 | 32 | 4 | 1 | 4 | 2.124s | 6.0 |
| 256 | 78B | 1 | 64 | 4 | 1 | 4 | 2.953s | 3.4 |
| 256 | 22B | 4 | 16 | 4 | 1 | 4 | 1.826s | 1.5 |
| | | | | | | | | |
32GB nodes:
| GPUs | Size | DP | PP | PP Chunks | Mic-BS | Glob-BS | Speed | TFlops |
| ---: | ---: | -: | -: | --------: | -----: | -----: | -----: | -----: |
| 16 | 18B | 1 | 4 | 4 | 1 | 4 | 1.381s | 26.7 |
| 32 | 30B | 1 | 8 | 4 | 1 | 4 | 1.618s | 19.0 |
| 64 | 65B | 1 | 16 | 4 | 1 | 4 | 2.738s | 12.2 |
| 128 | 116B | 1 | 32 | 4 | 1 | 4 | 4.234s | 7.0 |
| 256 | 206B | 1 | 64 | 4 | 1 | 4 | 6.736s | 3.9 |
| | | | | | | | | |
The TFLops are very low because there are too few PP chunks/micro-batches (4) (gradient accumulation size / GAS) and so the bubble takes a lot of overhead, increasing PP chunks should dramatically improve performance but also need to lower the max model size to have memory to hold those chunks in.
**Performance**
These experiments are to try a lower model size, but much higher TFlops performance
| GPUs | Size | DP | PP | PP Chunks | Mic-BS | Glob-BS | Speed | TFlops | Notes |
| ---: | ---: | -: | -: | --------: | -----: | -----: | ----: | -----: | ----: |
| 16 | 18B | 1 | 8 | 64 | 4 | 256 | 90.5s | 26.1 | 05-26 |
| 16 | 18B | 1 | 8 | 128 | 4 | 512 | 177s | 26.7 | 05-26 |
| 16 | 18B | 1 | 8 | 256 | 4 | 1024 | 356s | 26.5 | 05-26 |
| | | | | | | | | | |
| 16 | 18B | 1 | 4 | 128 | 4 | 512 | 179s | 26.4 | 05-26 |
| 16 | 18B | 1 | 4 | 128 | 6 | 768 | 262s | 27.0 | 05-26 |
| 16 | 18B | 1 | 8 | 128 | 6 | 768 | 259s | 27.3 | 05-26 |
| 16 | 18B | 1 | 8 | 32 | 8 | 256 | 89s | 26.5 | 05-26 |
| | | | | | | | | | |
| 32 | 39B | 1 | 8 | 128 | 4 | 512 | 82s | 62.3 | 05-26 |
| 32 | 39B | 1 | 8 | 128 | 6 | 768 | 123s | 62.3 | 05-26 |
| 32 | 39B | 1 | 8 | 256 | 6 | 1536 | 241s | 63.6 | 05-26 |
| 32 | 39B | 1 | 8 | 512 | 6 | 3072 | 478s | 64.2 | 05-26 |
| | | | | | | | | | |
| 64 | 52B | 1 | 16 | 256 | 4 | 1024 | 129s | 52.8 | 05-25 |
| 64 | 52B | 1 | 16 | 256 | 4 | 1024 | 217s | 31.4 | 05-26 |
| 64 | 52B | 1 | 16 | 256 | 4 | 1024 | 125s | 54.5 | 05-27 |
| 64 | 52B | 1 | 16 | 256 | 4 | 1024 | 225s | 30.3 | 05-28 |
| | | | | | | | | | |
| 64 | 52B | 1 | 16 | 256 | 6 | 1536 | 328s | 31.2 | 05-26 |
| 64 | 52B | 1 | 16 | 256 | 8 | 2048 | 435s | 31.3 | 05-26 |
| 64 | 52B | 1 | 16 | 512 | 6 | 3072 | 650s | 31.5 | 05-26 |
| 64 | 52B | 1 | 16 | 512 | 8 | 4096 | 870s | 31.3 | 05-26 |
| 64 | 52B | 1 | 32 | 256 | 4 | 1024 | 220s | 31.0 | 05-26 |
| | | | | | | | | | |
data:
- Size = Model Size
- `TP=4` in all of entries
- Speed is time per iteration - to complete global batch size
- Global batch size is `micro-batch-size * pp_chunks * dp_size`
- PP chunks is the number of PP stages, so each pipeline handles `micro-batch-size * pp_chunks`
- Seq length is 1024
notes:
- 32gpus had a very snag fit for gpu memory for 39B model (others were in ~75%) so it might be a bit too risky to OOM-borderline
#### Megatron + Deepspeed 3D (new branch)
Why:
1. More generic pipeline API that is not hard-coded into the model
2. Better memory efficiency - needs less GPU memory, so can probably work with fewer pipeline stages
3. Works with ZeRO-Offload so can significantly reduce the GPUs required for fine-tuning once the model is pre-trained, making it accessible to a lot more folks, who don't have access to hundreds of GPUs.
How:
This is new branch synced with Megatron
DeepSpeed branch: https://github.com/ShadenSmith/DeepSpeed/tree/megatron2.4-3d
Megatron branch: https://github.com/jeffra/DSE/tree/megatron-2.4-ds-pipe
This script can now launch Meg alone or Meg + Deepspeed 3D (ignore the zero options it doesn't work yet):
https://github.com/jeffra/DSE/blob/megatron-2.4-ds-pipe/run.sh
```
git clone https://github.com/ShadenSmith/DeepSpeed/ deepspeed-shaden
cd deepspeed-shaden
git checkout megatron2.4-3d
```
```
git clone https://github.com/jeffra/DSE megator-jeffra
cd megator-jeffra
git checkout megatron-2.4-ds-pipe
```
See scripts and logs under [gpt2-meg-ds-3d](./gpt2-meg-ds-3d).
Now we use the same code-base for training w/ and w/o DS/3D - so can use a shared results table.
Also added memory usage columns.
| GPUs | Size | DS | GPU M | DP | PP | GAS | MBS | GBS | Speed | TFlops | Notes |
| ---: | ---: | -: | ----: | -: | -: | ---: | --: | ---: | ----: | -----: | ----: |
| 64 | 52B | Y | 26GB | 1 | 16 | 256 | 4 | 1024 | 137s | 46.7 | 06-10 |
| 64 | 52B | Y | 29GB | 1 | 16 | 256 | 4 | 1536 | 206s | 49.6 | 06-10 |
| 64 | 52B | Y | 32GB | 1 | 16 | 256 | 4 | 2048 | 270s | 50.5 | 06-10 |
| 64 | 52B | Y | 26GB | 1 | 16 | 1024 | 4 | 4096 | 544s | 50.1 | 06-10 |
| | | | | | | | | | | | |
| | | | | | | | | | | | |
| 64 | 52B | N | 32GB | 1 | 16 | 256 | 4 | 1024 | 126s | 54.1 | 06-10 |
| | | | | | | | | | | | |
```
perl -le '$ng=64; $ms=52; $gbs=1024; $sp=146; print $ms*4*2*1024*$gbs / ( $sp * $ng * 1e3)'
```
- DS: Deepspeed/3D enabled
- GPU memory: rounded up per GPU
- MBS: Micro BS
- GBS: Global BS = GAS * MBS * DP_SIZE
- GAS: Gradient Accumulation Steps (= MBS pipe stages, = PP chunks)
Resident CPU memory remained at about 3GB per GPU.
**zero_stage:1 + reduce_bucket_size**
also added `--partition-activations`
(`meg_ds_3d_gpt2_perf_n16_z1_try*.slurm`)
| GPUs | Size | DS | bucket | DP | PP | GAS | MBS | GBS | Speed | TFlops | Notes |
| ---: | ---: | -: | ----: | -: | -: | ---: | --: | ---: | ----: | -----: | ----: |
| 64 | 52B | Y | 5e8 | 2 | 8 | 128 | 4 | 1024 | 137s | 48.8 | 07-10 |
| 64 | 52B | Y | 1e9 | 2 | 8 | 128 | 4 | 1024 | 141s | 48.3 | 07-10 |
| 64 | 52B | Y | 2e9 | 2 | 8 | 128 | 4 | 1024 | 141s | 48.3 | 07-10 |
| | | | | | | | | | | | |
Note: Since PP*TP=8*4=32, so since there are 64GPUs - DP=2
------------
Experiment 1:
TP=4, DP=2, PP=8, gas=256, DS_ZeRO Stage 1, PA=disabled,reduce_bucket_size=2e8,5e8, mbs=2,3,
| ID | GPUs | Size | DS | bucket | DP | PP | GAS | MBS | GBS | Speed | TFlops | Notes |
| --: | ---: | ---: | -: | ----: | -: | -: | ---: | --: | ---: | ----: | -----: | ----: |
| 1.1 | 64 | 52B | Y | 2e8 | 2 | 8 | 256 | 2 | 1024 | 150s | 45.4 | 07-10 |
| 1.2 | 64 | 52B | Y | 5e8 | 2 | 8 | 256 | 2 | 1024 | 150s | 45.4 | 07-10 |
| 1.3 | 64 | 52B | Y | 2e8 | 2 | 8 | 256 | 3 | 1536 | 213 | 48.0 | 07-10 |
| 1.4 | 64 | 52B | Y | 5e8 | 2 | 8 | 256 | 3 | 1536 | 208 | 49.1 | 07-10 |
| | | | | | | | | | | | | |
------------
Experiment 2: HD=8192, NUM_LAYERs=48 (MSIZE=39)
Megatron+DeepSpeed:
- USE_DEEPSPEED=1, MSIZE=39, TP=4, PP=8, DP=2, ZeRO Stage 1, mbs=4, PA=disabled, reduce_bucket_size=2e8, gas=128
- USE_DEEPSPEED=1, MSIZE=39, TP=4, PP=8, DP=2, ZeRO Stage 1, mbs=4, PA=disabled, reduce_bucket_size=5e8, gas=128
Megatron Alone (which ever of the following runs better)
- USE_DEEPSPEED=0, MSIZE=39, TP=4, PP=16, DP=1, mbs=4, gas=256
- USE_DEEPSPEED=0, MSIZE=39, TP=4, PP =8, DP=2, mbs=4, gas=128
| ID | GPUs | Size | DS | bucket | DP | PP | GAS | MBS | GBS | Speed | TFlops | Notes |
| --: | ---: | ---: | -: | ----: | -: | -: | ---: | --: | ---: | ----: | -----: | ----: |
| 2.1 | 64 | 39B | Y | 2e8 | 2 | 8 | 128 | 4 | 1024 | 104s | 49.1 | 07-10 |
| 2.2 | 64 | 39B | Y | 5e8 | 2 | 8 | 128 | 4 | 1024 | 105s | 48.7 | 07-10 |
| 2.3 | 64 | 39B | N | na | 1 | 8 | 256 | 4 | 1024 | 109s | 46.9 | 07-10 |
| 2.4 | 64 | 39B | N | na | 2 | 8 | 128 | 4 | 1024 | 110s | 46.5 | 07-10 |
| | | | | | | | | | | | | |
------------
note: I also did tests on 1 node - getting almost identical results for Meg w/ and w/o DS/3D. So all the fluctuations are the network to blame for.
```
NNODES=1
PP_SIZE=1
TP_SIZE=4
MICRO_BATCH_SIZE=4
PP_CHUNKS=16 # GAS
MSIZE=4
```
got an average over 22 iterations in msecs (too short for good stats)
```
ds 6875.05
meg 6896.20
```
but it's obvious they are pretty similar.
**save-checkpoint speed Measurement**
| Nodes | MSize | Time (ms) |
| ----: | ----: | -------: |
| 8 | 25B | 17960.68 |
| | | |
| 16 | 52B | 19298.14 |
| 32 | 52B | 19228.38 |
| 64 | 52B | 19652.80 |
| | | |
| 32 | 97B | 19417.09 |
| 64 | 97B | 11525.99 |
| | | |
| 64 | 181B | 19495.31 |
| | | |
Currently it saves everything, not just model weights.
The biggest test was for 181B model, 64 nodes, 256 gpus, and a total 2.4TB per checkpoint.
The breakdown is:
1. 0.34TB in PP layer states, 1.4GB per file per gpu (1.4*256) - this one looks like 2bytes per param
2. 2.00TB in optimizer states, 8.0GB per file per gpu (8*256) - this one looks like 12bytes per param
The data sizes are correspondingly:
1. 2 bytes per param for fp16 weights
2. 12 bytes are 8 bytes for optimizer and 4 bytes for fp32 model
To make lots of these we should copy away only the fp16 weights, and overwrite the checkpoint - otherwise a lot more HD space will be needed.
Important: also remember that `$six_ALL_CCFRSCRATCH` files that don't get accessed in 30 days get auto-deleted, so the important checkpoints need to be backed up (probably tar'ed and put on `$six_ALL_CCFRSTORE`).
### Megatron + Deepspeed 3D (old branch)
**Important**: `DeepSpeedExamples/Megatron-LM-v1.1.5-3D_parallelism` is not in sync with M-LM master - so several config args don't match. It's about 8 months old.
See scripts and logs under [gpt2-meg-ds-3d-old](./gpt2-meg-ds-3d-old).
Uses 3D:
- TP: tensor parallelism
- PP: pipeline parallelism
- DP: data parallelism
same features as Megatron's native, but improved by Deepspeed
**Performance**
| GPUs | Size | DP | PP | PP chunks | Mic-BS | Glob-BS | Speed | TFlops | Notes |
| ---: | ---: | -: | -: | --------: | -----: | ------: | ----: | -----: | ----: |
| 64 | 52B | 1 | 16 | 256 | 4 | 1024 | 146s | 46.7 | 05-27 |
| | | | | | | | | | |
- GAS = Gradient Accumulation size (same as PP_chunks / number of PP stages)
- Global_bs = pp_chunks*micro_bs*dp_size
- `TP_SIZE=4` (size of the node)
```
perl -le '$ng=64; $ms=52; $gbs=1024; $sp=146; print $ms*4*2*1024*$gbs / ( $sp * $ng * 1e3)'
```
#### Megatron + Deepspeed ZeRO (old branch)
**Important**: `DeepSpeedExamples/Megatron-LM-v1.1.5-ZeRO3` is not in sync with M-LM master - so several config args don't match. It's about 8 months old.
See scripts and logs under [gpt2-meg-ds-zero](./gpt2-meg-ds-zero).
This one uses only TP from Megatron (no PP)
Not yet optimized with Deepspeed team!
**With Offload off**
**Performance**
| GPUs | Size | DP | Mic-BS | Glob-BS | Speed | TFlops | Notes |
| ---: | ----: | -: | ---: | -----: | ----: | -----: | ----: |
| 64 | 52B | 16 | 48 | 768 | 122s | 41.9 | 05-25 |
| 64 | 52B | 16 | 48 | 768 | 127s | 40.3 | 05-27 |
| | | | | | | | |
```
perl -le '$ng=64; $ms=52; $gbs=768; $sp=122; print $ms*4*2*1024*$gbs / ( $sp * $ng * 1e3)'
```
- Seq length is 1024
- `TP=4` in all of entries
- `DP` is number of nodes here
- Speed is time per iteration - to complete global batch size
- Global batch size is `micro-batch-size * dp-size`
- tried w/ and w/o Tiling once but saw no difference - perhaps would be more important on larger collections
| GPUs | Size | TP | DP | Mic-BS | Glob-BS | Speed | TFlops | Notes |
| ---: | ---: | -: | -: | -----: | ------: | ----: | -----: | ----: |
| 64 | 52B | 4 | 16 | 48 | 768 | 127s | 40.3 | |
| 64 | 52B | 2 | 32 | 32 | 1024 | 167s | 40.8 | |
| 64 | 52B | 1 | 64 | 16 | 1024 | 184s | 37.0 | |
| 64 | 24B | 1 | 64 | 16 | 1024 | 89.0s | 35.3 | |
| 64 | 24B | 2 | 32 | 32 | 1024 | 85.7s | 36.7 | |
**With full cpu offload**
| GPUs | Size | TP | DP | Mic-BS | Glob-BS | Speed | TFlops |
| ---: | ---: | -: | -: | -----: | ------: | ----: | -----: |
| 64 | 52B | 4 | 16 | 64 | 1024 | 171s | 39.9 |
| | | | | | | | |
Olatunji requested the following experiments:
- enabled/set: `--split-transformers --checkpoint-num-layers=2`
- removed: `--synchronize-each-layer --contigious-checkpointing`
| ID | GPUs | Size | ScatEmb | TP | DP | Mic-BS | Glob-BS | Speed | TFlops |
| -: | ---: | ---: | ------: | -: | -: | -----: | ------: | ----: | -----: |
| 1 | 64 | 52B | N | 4 | 16 | 48 | 768 | 119s | 43.0 |
| 2 | 64 | 52B | Y | 4 | 16 | 48 | 768 | 115s | 44.5 |
| 3 | 64 | 52B | Y | 4 | 16 | 52 | 832 | 124s | 44.7 |
| 4 | 64 | 52B | N | 2 | 32 | 32 | 1024 | 159s | 42.9 |
| 5 | 64 | 52B | Y | 2 | 32 | 32 | 1024 | 158s | 43.1 |
| 6 | 64 | 52B | Y | 2 | 32 | 36 | 1152 | 176s | 43.6 |
| 7 | 64 | 52B | Y | 4 | 16 | 56 | 896 | 161s | 37.0 |
| 8 | 64 | 52B | Y | 2 | 32 | 38 | 1216 | 178s | 45.5 |
| 9 | 64 | 52B | Y | 1 | 64 | 18 | 1152 | 197s | 38.9 |
| 10 | 64 | 52B | Y | 1 | 64 | 20 | 1280 | 219s | 38.9 |
| 11 | 64 | 52B | Y | 1 | 64 | 22 | 1408 | OOM | |
| | | | | | | | | | |
following 2:
from ID 8:
- removed `--checkpoint-in-cpu'
- changed values
| ID | GPUs | Size | ScatEmb | TP | DP | Mic-BS | Glob-BS | Speed | TFlops |
| -: | ---: | ---: | ------: | -: | -: | -----: | ------: | ----: | -----: |
| 12 | 64 | 52B | Y | 4 | 16 | 24 | 384 | 72s | 35.5 |
| 13 | 64 | 52B | Y | 2 | 32 | 16 | 512 | 79s | 38.3 |
| | | | | | | | | | |
following 4:
from ID 12:
- removed `--split-transformers`
- changed values
- toggled `--checkpoint-in-cpu` (PA_CPU column)
| ID | GPUs | Size | ScatEmb | PA_CPU | TP | DP | Mic-BS | Glob-BS | Speed | TFlops |
| -: | ---: | ---: | ------: | -----: | -: | -: | -----: | ------: | ----: | -----: |
| 14 | 64 | 52B | Y | N | 4 | 16 | 24 | 384 | 72s | 35.5 |
| 15 | 64 | 52B | Y | Y | 4 | 16 | 24 | 384 | 71s | 36.0 |
| 16 | 64 | 52B | Y | N | 2 | 32 | 16 | 512 | 87s | 39.2 |
| 17 | 64 | 52B | Y | Y | 2 | 32 | 16 | 512 | 88s | 38.7 |
| | | | | | | | | | | |
### HF + Deepspeed Zero 3 + Full Offload
See scripts and logs under [gpt2-hf-ds](./gpt2-hf-ds).
Not yet optimized with Deepspeed team!
**Max model size**
| GPUs | Size | Mic-BS | Glob-BS | Speed | TFlops |
| ---: | ----: | -----: | ------: | ----: | -----: |
| 16 | 25B | 4 | 64 | 58s | 14.0 |
| 32 | 52B | 4 | 128 | 114s | 14.9 |
| 64 | 97B | 4 | 256 | 222s | 14.3 |
| | | | | | |
**Performance**
| GPUs | Size | Zero | Opt Offl | Par Offl | Mic-BS | Glob-BS | Speed | TFlops | Notes |
| ---: | ----: | --: | -------: | -------: | -----: | ------: | ----: | -----: | ----: |
| 64 | 52B | 3 | N | N | 8 | 512 | 139s | 24.5 | 05-25 |
| 64 | 52B | 3 | N | N | 4 | 256 | 185s | 9.2 | 05-27 |
| 64 | 52B | 3 | N | N | 8 | 512 | 118s | 28.9 | 05-27 |
| | | | | | | | | | |
| 64 | 52B | 3 | N | N | 8 | 512 | 117s | 29.1 | 05-28 |
| 64 | 52B | 3 | N | N | 6 | 384 | 111s | 23.0 | 05-28 |
| 64 | 52B | 3 | N | N | 10 | 640 | 150s | 28.4 | 05-28 |
| 64 | 52B | 3 | Y | N | 12 | 768 | 183s | 27.9 | 05-28 |
| 64 | 52B | 3 | Y | N | 12 | 768 | 175s | 29.2 | 05-28 |
| 64 | 52B | 3 | Y | Y | 12 | 768 | 177s | 28.9 | 05-28 |
| | | | | | | | | | |
| 64 | 52B | 2 | Y | N | | | OOM | | 05-28 |
| | | | | | | | | | |
- DP=GPUs
- global bs = micro bs * DP
- Speed reported by HF Trainer metrics is `samples_per_second` - So total throughput in the table is `glob_bs/samples_per_second`
notes:
- gradient checkpointing activated
```
perl -le '$ng=64; $ms=52; $gbs=512; $sp=139.52; print $ms*4*2*1024*$gbs / ( $sp * $ng * 1e3)'
22
```
ZeRO-2 with model of this size I can't fit into this setup at all - even BS=4 - it keeps getting on getting killed by cgroups - i.e. it's asking for more than 40GB general RAM per gpu. Same story w/ or w/o offload.
## Magic scripts
- Calculate the TFlops:
```
perl -le '$ng=64; $ms=52; $gbs=1024; $sp=127; print $ms*4*2*1024*$gbs / ( $sp * $ng * 1e3)'
```
(ng = total gpus, ms = model size in B, gbs = global batch size, sp = throughput in seconds)
same with bash env vars and broken down GBS into mbs*dp*gas (gas=pp_chunks):
```
echo "($MSIZE*4*2*1024*$MICRO_BATCH_SIZE*$DP_SIZE*$PP_CHUNKS)/($THROUGHPUT*$NNODES*4*1000)" | bc -l
```
- Automatically process slurm/ megatron log files, average the throughput (prints 'fail' on when the training failed w/o producing a single iteration stat):
```
find . -type f -name "*out" -exec perl -lne 'm|elapsed time per iteration .ms.: ([\d\.]+)| && do {$x+=$1; $c++}; END { print "$ARGV " . ($c ? int($x/$c/1000) : "fail")}' {} \; | sort | grep -v fail
```
- re-generate tflops column in the tables above:
```
perl -ne 's#^(\| +(\d+) +\| +(\d+)B.*? +(\d+) +\| +([\d\.]+)s) +\| +[\d\.]+ +(.*?)$#"$1 | ".sprintf("%.01f", $3*4*2*1024*$4 / ($5*$2*1e3))." $6"#e && print ' gpt2.md
```
I originally had a mistake in model size calculation script - which has been fixed in tables and the scripts, but many logs still have the old formula - I used G `(2**30)` instead of B `(10**9)` so the model size was getting reported smaller than it is.
Now it's the correct version:
```
NHIDDEN=4096
NLAYERS=36
SEQ_LEN=512
VOCAB_SIZE=50257
python -c "h=$NHIDDEN; l=$NLAYERS; s=$SEQ_LEN; v=$VOCAB_SIZE; print(f'Model size: {(l*(12*h**2 + 13*h) + v*h + s*h + 2*h) / 10**9 :.0f}B')"
```
- misc file renames
```
# rename both .sh and .out based on GAS (PP_CHUNKS) value inside
# 61B-megatron-mbs-2-pp16-dp-1.sh -> 61B-megatron-mbs-2-pp16-dp-1-gas128.sh
perl -lne 'm|PP_CHUNKS=(\d+)| && do {$gas=$1; $q = chr(39); $ARGV=~s|\.sh$||; print qq[rename.pl ${q}s|dp-(\\d)|dp-\$1-gas-$gas|$q $ARGV*] }' *sh > run-renames.sh
sh ./run-renames.sh
```
- A formula to match the script name to the log file, by rewriting the `job-name`:
```
perl -pi -e '$ARGV=~s|\.sh$||; s|#SBATCH --job-name=.*|#SBATCH --job-name=$ARGV|' *slurm *sh
```
now the log file will match the slurm file.
- change runtime:
```
perl -pi -e '$ARGV=~s|\.sh$||; s|#SBATCH --time=.*|#SBATCH --time=00:20:00|' *slurm *sh
```
- calculate speed + tflops from filename and averaging `elapsed time per iteration` from the log - including failed runs (needs the `-gas-` file rename from above)
```
find . -type f -name "*out" -exec perl -lne 'm|elapsed time per iteration .ms.: ([\d\.]+)| && do {$x+=$1; $c++}; END { $sp=$c ? int($x/$c/1000) : 0; $d=qr/(\d+)/; $ARGV=~m|${d}B-.*?-mbs-$d-pp$d-dp-$d-gas-$d| && do {$ng=64; $ms=$1; $gbs=$2*$4*$5; $tf=$sp ? sprintf "%0.1f", $ms*4*2*1024*$gbs / ( $sp * $ng * 1e3) : 0}; $r = $sp ? "$ARGV $sp $tf" : "$ARGV fail"; print $r}' {} \; | sort -nk3 -r
./61B-megatron-mbs-2-pp16-dp-1-gas-512-200977.out 144 55.5
./55B-megatron-mbs-2-pp16-dp-1-gas-512-200968.out 134 53.8
./55B-ds-zero0-mbs-2-pp16-dp-1-gas-512-200964.out 141 51.1
./55B-ds-zero0-mbs-4-pp16-dp-1-gas-256-200965.out 145 49.7
./55B-megatron-mbs-4-pp16-dp-1-gas-256-200970.out 149 48.4
./61B-ds-zero0-mbs-4-pp16-dp-1-gas-256-200973.out 166 48.2
./61B-ds-zero0-mbs-2-pp16-dp-1-gas-512-200972.out 169 47.3
./61B-megatron-mbs-4-pp16-dp-1-gas-256-200979.out 172 46.5
./61B-megatron-mbs-4-pp8-dp-2-gas-128-200980.out fail
./61B-megatron-mbs-2-pp8-dp-2-gas-256-200978.out fail
./61B-ds-zero1-mbs-4-pp8-dp-2-gas-128-200976.out fail
./61B-ds-zero1-mbs-2-pp8-dp-2-gas-256-200974.out fail
./55B-megatron-mbs-4-pp8-dp-2-gas-128-200971.out fail
./55B-megatron-mbs-2-pp8-dp-2-gas-256-200969.out fail
./55B-ds-zero1-mbs-4-pp8-dp-2-gas-128-200967.out fail
./55B-ds-zero1-mbs-2-pp8-dp-2-gas-256-200966.out fail
```
- same as above but with finer control over which files are processed and preserving their run order, e.g. sorted by latest run:
```
ls -1t 61*out | xargs -n1 perl -lne 'm|elapsed time per iteration .ms.: ([\d\.]+)| && do {$x+=$1; $c++}; END { $sp=$c ? int($x/$c/1000) : 0; $d=qr/(\d+)/; $ARGV=~m|${d}B-.*?-mbs-$d-pp$d-dp-$d-gas-$d| && do {$ng=64; $ms=$1; $gbs=$2*$4*$5; $tf=$sp ? sprintf "%0.1f", $ms*4*2*1024*$gbs / ( $sp * $ng * 1e3) : 0}; $r = $sp ? "$ARGV $sp $tf" : "$ARGV fail"; print $r}'
61B-ds-zero1-mbs-2-pp16-dp-1-gas-512.18488.out 196 40.8
61B-megatron-mbs-2-pp16-dp-1-gas-512.8189.out 176 45.4
61B-ds-zero1-mbs-2-pp16-dp-1-gas-512.17709.out 194 41.2
```
|