| # Train 8 104B wide tune up | |
| note: this tune up table is somewhat invalid since during the tune up a mistake was made in `FFN_HIDDEN_SIZE` which was incorrectly set to a much lower value, so the tests below were really tested a 58B model. So the TFLOPs numbers in this section are incorrect (bigger than they are in reality). but I'm not sure how to fix it, since I don't think the formula applies when the model is lopsided. The numbers in sections afterwards are correct. | |
| The misconfiguration error has been fixed later in the experiments. | |
| ``` | |
| NLAYERS=32 | |
| NHIDDEN=16384 | |
| NHEADS=32 | |
| SEQ_LEN=2048 | |
| VOCAB_SIZE=50257 | |
| ``` | |
| BS=1024, SIZE=104B, | |
| | NNODES | TP | PP | DP | MBS | Speed | TFlops | Notes | | |
| | -----: | --: | --: | --: | --: | ----: | -----: | --------------------: | | |
| | 32 | 4 | 32 | 1 | 1 | 256 | 54.5 | 31.5GB | | |
| | 64 | 4 | 64 | 1 | 1 | 155 | 55.0 | 24GB | | |
| | | | | | | | | | | |
| ``` | |
| perl -le '$ng=32*4; $sp=256; $ms=104; $gbs=1048; $seqlen=2048; print $ms*4*2*$seqlen*$gbs / ( $sp * $ng * 1e3)' | |
| perl -le '$ng=64*4; $sp=155; $ms=104; $gbs=1048; $seqlen=2048; print $ms*4*2*$seqlen*$gbs / ( $sp * $ng * 1e3)' | |
| ``` | |
| (ng = total gpus, ms = model size in B, gbs = global batch size, sp = throughput in seconds) | |
| BS=2048 | |
| | NNODES | TP | PP | DP | MBS | Speed | TFlops | Notes | | |
| | ----: | --: | --: | --: | --: | ----: | -----: | --------------------: | | |
| | 32 | 4 | 32 | 1 | 1 | 586 | 46.52 | GB | | |
| | 64 | 4 | 64 | 1 | 1 | 301 | 45.28 | 25GB | | |
| | | | | | | | | | | |
| ``` | |
| perl -le '$ng=32*4; $sp=586; $ms=104; $gbs=2048; $seqlen=2048; print $ms*4*2*$seqlen*$gbs / ( $sp * $ng * 1e3)' | |
| perl -le '$ng=64*4; $sp=301; $ms=104; $gbs=2048; $seqlen=2048; print $ms*4*2*$seqlen*$gbs / ( $sp * $ng * 1e3)' | |
| ``` | |
| e.g. interactive tuning on 32 nodes | |
| ``` | |
| salloc --account=six@gpu --constraint=v100-32g --nodes=32 --ntasks=32 --cpus-per-task=40 --gres=gpu:4 --hint=nomultithread --time=3:00:00 bash --rcfile $six_ALL_CCFRWORK/start-prod | |
| ``` | |
| ## BNB | |
| w/ `--use-bnb-optimizer` | |
| | NNODES | TP | PP | DP | MBS | Speed | TFlops | Notes | | |
| | ----: | --: | --: | --: | --: | ----: | -----: | --------------------: | | |
| | 32 | 4 | 16 | 2 | 1 | 681 | 40.0 | 31GB | | |
| | 32 | 2 | 32 | 2 | 1 | 633 | 43.0 | 31GB | | |
| | 32 | 1 | 64 | 2 | 1 | | | 32GB OOMs | | |
| | 32 | 4 | 32 | 1 | 1 | 688 | 39.6 | 27GB (same conf as normal 104B) | | |
| | | | | | | | | | | |
| ``` | |
| perl -le '$ng=32*4; $sp=633; $ms=104; $gbs=2048; $seqlen=2048; print $ms*4*2*$seqlen*$gbs / ( $sp * $ng * 1e3)' | |
| ``` | |
| To ensure we are comparing apples to apples, trying to using the same allocations re-testing the baseline (but I'm not I get the same nodes all the time). | |
| The baseline of 104B experiment w/o `--use-bnb-optimizer` that we have been using for all experiments | |
| using the `main` branch: | |
| | NNODES | TP | PP | DP | MBS | Speed | TFlops | Notes | | |
| | ----: | --: | --: | --: | --: | ----: | -----: | --------------------: | | |
| | 32 | 4 | 32 | 1 | 1 | 696 | 39.17 | 30GB (same conf as normal 104B) | | |
| | | | | | | | | | | |
| using the old `big-science` branch | |
| | NNODES | TP | PP | DP | MBS | Speed | TFlops | Notes | | |
| | ----: | --: | --: | --: | --: | ----: | -----: | --------------------: | | |
| | 32 | 4 | 32 | 1 | 1 | 706 | 38.6 | 30GB (same conf as normal 104B) | | |
| | | | | | | | | | | |
| ## A100s | |
| GPUS_PER_NODE=8 | |
| NNODES=16 | |
| TP_SIZE=4 # always fixed to the size of a single node | |
| PP_SIZE=32 # NLAYERS must be a multiple of PP_SIZE here | |
| MICRO_BATCH_SIZE=1 | |
| GLOBAL_BATCH_SIZE=2048 | |
| TFLOPs: 72.72-82 (was speeding up - so very inconclusive) | |