Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "432a4f4d45e9eb6c3e31971d1e0e69a9bc852a21"
Commit 5e56e563 authored by Neel Kant's avatar Neel Kant
Browse files

Merge master into realm-mlm

parents 6c0a5bd8 569b3dab
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 752 additions and 624 deletions
LICENSE
View file @ 5e56e563
# Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved. The following applies to all files unless otherwise noted:
# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
# #
# Redistribution and use in source and binary forms, with or without # Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions # modification, are permitted provided that the following conditions
...@@ -24,8 +26,15 @@ ...@@ -24,8 +26,15 @@
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
--
This repository also contains code from Hugging Face Inc., Google Research,
and Facebook (from their Fairseq project). Files from these
organizations have notices at the top of each file. Below are licenses
used in those files, as indicated.
------------- LICENSE FOR huggingface(transformer) repository -------------- ------------- LICENSE FOR huggingface and Google Research code --------------
Apache License Apache License
...@@ -229,3 +238,27 @@ ...@@ -229,3 +238,27 @@
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and See the License for the specific language governing permissions and
limitations under the License. limitations under the License.
------------- LICENSE FOR Facebook Fairseq code --------------
MIT License
Copyright (c) Facebook, Inc. and its affiliates.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
README.md
View file @ 5e56e563
This diff is collapsed.
examples/evaluate_zeroshot_gpt2.sh 0 → 100755
View file @ 5e56e563
#!/bin/bash
WORLD_SIZE=8
DISTRIBUTED_ARGS="--nproc_per_node $WORLD_SIZE \
--nnodes 1 \
--node_rank 0 \
--master_addr localhost \
--master_port 6000"
TASK="LAMBADA"
VALID_DATA=<lambada path>
VOCAB_FILE=gpt2-vocab.json
MERGE_FILE=gpt2-merges.txt
CHECKPOINT=checkpoints/gpt2_345m
python -m torch.distributed.launch $DISTRIBUTED_ARGS ./tasks/main.py \
--task $TASK \
--valid-data $VALID_DATA \
--tokenizer-type GPT2BPETokenizer \
--strict-lambada \
--vocab-file $VOCAB_FILE \
--merge-file $MERGE_FILE \
--load $CHECKPOINT \
--model-parallel-size 1 \
--num-layers 24 \
--hidden-size 1024 \
--num-attention-heads 16 \
--batch-size 8 \
--checkpoint-activations \
--seq-length 1024 \
--max-position-embeddings 1024 \
--log-interval 10 \
--fp16 \
--no-load-optim \
--no-load-rng
examples/finetune_mnli_distributed.sh 0 → 100755
View file @ 5e56e563
#!/bin/bash
WORLD_SIZE=8
DISTRIBUTED_ARGS="--nproc_per_node $WORLD_SIZE \
--nnodes 1 \
--node_rank 0 \
--master_addr localhost \
--master_port 6000"
TRAIN_DATA="data/glue_data/MNLI/train.tsv"
VALID_DATA="data/glue_data/MNLI/dev_matched.tsv \
data/glue_data/MNLI/dev_mismatched.tsv"
PRETRAINED_CHECKPOINT=checkpoints/bert_345m
VOCAB_FILE=bert-vocab.txt
CHECKPOINT_PATH=checkpoints/bert_345m_mnli
python -m torch.distributed.launch $DISTRIBUTED_ARGS ./tasks/main.py \
--task MNLI \
--seed 1234 \
--train-data $TRAIN_DATA \
--valid-data $VALID_DATA \
--tokenizer-type BertWordPieceLowerCase \
--vocab-file $VOCAB_FILE \
--epochs 5 \
--pretrained-checkpoint $PRETRAINED_CHECKPOINT \
--model-parallel-size 1 \
--num-layers 24 \
--hidden-size 1024 \
--num-attention-heads 16 \
--batch-size 8 \
--checkpoint-activations \
--lr 5.0e-5 \
--lr-decay-style linear \
--warmup 0.065 \
--seq-length 512 \
--max-position-embeddings 512 \
--save-interval 500000 \
--save $CHECKPOINT_PATH \
--log-interval 10 \
--eval-interval 100 \
--eval-iters 50 \
--weight-decay 1.0e-1 \
--fp16
examples/finetune_race_distributed.sh 0 → 100755
View file @ 5e56e563
#!/bin/bash
WORLD_SIZE=8
DISTRIBUTED_ARGS="--nproc_per_node $WORLD_SIZE \
--nnodes 1 \
--node_rank 0 \
--master_addr localhost \
--master_port 6000"
TRAIN_DATA="data/RACE/train/middle"
VALID_DATA="data/RACE/dev/middle \
data/RACE/dev/high"
VOCAB_FILE=bert-vocab.txt
PRETRAINED_CHECKPOINT=checkpoints/bert_345m
CHECKPOINT_PATH=checkpoints/bert_345m_race
python -m torch.distributed.launch $DISTRIBUTED_ARGS ./tasks/main.py \
--task RACE \
--seed 1234 \
--train-data $TRAIN_DATA \
--valid-data $VALID_DATA \
--tokenizer-type BertWordPieceLowerCase \
--vocab-file $VOCAB_FILE \
--epochs 3 \
--pretrained-checkpoint $PRETRAINED_CHECKPOINT \
--model-parallel-size 1 \
--num-layers 24 \
--hidden-size 1024 \
--num-attention-heads 16 \
--batch-size 4 \
--checkpoint-activations \
--lr 1.0e-5 \
--lr-decay-style linear \
--warmup 0.06 \
--seq-length 512 \
--max-position-embeddings 512 \
--save-interval 100000 \
--save $CHECKPOINT_PATH \
--log-interval 10 \
--eval-interval 100 \
--eval-iters 50 \
--weight-decay 1.0e-1 \
--clip-grad 1.0 \
--hidden-dropout 0.1 \
--attention-dropout 0.1 \
--fp16
examples/generate_text.sh
View file @ 5e56e563
#!/bin/bash #!/bin/bash
CHECKPOINT_PATH=checkpoints/gpt2_345m/ CHECKPOINT_PATH=checkpoints/gpt2_345m
MPSIZE=1 VOCAB_FILE=gpt2-vocab.json
NLAYERS=12 MERGE_FILE=gpt2-merges.txt
NHIDDEN=768
NATT=12
MAXSEQLEN=1024
#SAMPLING ARGS python tools/generate_samples_gpt2.py \
TEMP=0.9 --model-parallel-size 1 \
#If TOPK/TOPP are 0 it defaults to greedy sampling, top-k will also override top-p --num-layers 24 \
TOPK=0 --hidden-size 1024 \
TOPP=0
python generate_samples.py \
--model-parallel-size $MPSIZE \
--num-layers $NLAYERS \
--hidden-size $NHIDDEN \
--load $CHECKPOINT_PATH \ --load $CHECKPOINT_PATH \
--num-attention-heads $NATT \ --num-attention-heads 16 \
--max-position-embeddings 1024 \ --max-position-embeddings 1024 \
--tokenizer-type GPT2BPETokenizer \ --tokenizer-type GPT2BPETokenizer \
--fp16 \ --fp16 \
--cache-dir cache \ --batch-size 2 \
--out-seq-length $MAXSEQLEN \ --seq-length 1024 \
--temperature $TEMP \ --out-seq-length 1024 \
--top_k $TOPK \ --temperature 1.0 \
--genfile dbg_unconditional.json \ --vocab-file $VOCAB_FILE \
--num-samples 10 \ --merge-file $MERGE_FILE \
--top_p $TOPP \ --genfile unconditional_samples.json \
--num-samples 2 \
--top_p 0.9 \
--recompute --recompute
examples/merge_mp_bert.sh 0 → 100755
View file @ 5e56e563
#!/bin/bash
MODEL_PARALLEL_SIZE=2
VOCAB_FILE=bert-vocab.txt
CHECKPOINT_PATH=checkpoints/bert_345m
WORLD_SIZE=$MODEL_PARALLEL_SIZE python tools/merge_mp_partitions.py \
--model-type BERT \
--model-parallel-size $MODEL_PARALLEL_SIZE \
--tokenizer-type BertWordPieceLowerCase \
--vocab-file $VOCAB_FILE \
--num-layers 24 \
--hidden-size 1024 \
--num-attention-heads 16 \
--seq-length 512 \
--max-position-embeddings 512 \
--load $CHECKPOINT_PATH
examples/pretrain_albert.sh deleted 100755 → 0
View file @ 6c0a5bd8
#!/bin/bash
RANK=0
WORLD_SIZE=1
python pretrain_albert.py \
--num-layers 12 \
--hidden-size 768 \
--num-attention-heads 12 \
--batch-size 4 \
--seq-length 512 \
--max-preds-per-seq 80 \
--max-position-embeddings 512 \
--train-iters 10000 \
--save checkpoints/albert_117m \
--load checkpoints/albert_117m \
--resume-dataloader \
--data-path data/megatron/bc_rn_owt_sto_wiki_dedup_shuf_cleaned_0.7_mmap \
--vocab data/megatron/vocab.txt \
--split 949,50,1 \
--distributed-backend nccl \
--lr 0.0001 \
--lr-decay-style linear \
--lr-decay-iters 990000 \
--weight-decay 1e-2 \
--clip-grad 1.0 \
--warmup .01 \
--fp16 \
--fp32-layernorm \
--fp32-embedding \
--skip-mmap-warmup \
--num-workers 0
examples/pretrain_albert_distributed.sh deleted 100755 → 0
View file @ 6c0a5bd8
#!/bin/bash
GPUS_PER_NODE=2
# Change for multinode config
MASTER_ADDR=localhost
MASTER_PORT=6000
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES))
DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT"
python -m torch.distributed.launch $DISTRIBUTED_ARGS \
pretrain_albert.py \
--num-layers 12 \
--hidden-size 768 \
--num-attention-heads 12 \
--batch-size 4 \
--seq-length 512 \
--max-preds-per-seq 80 \
--max-position-embeddings 512 \
--train-iters 10000 \
--save checkpoints/albert_117m \
--load checkpoints/albert_117m \
--resume-dataloader \
--data-path data/megatron/bc_rn_owt_sto_wiki_dedup_shuf_cleaned_0.7_mmap \
--vocab data/megatron/vocab.txt \
--split 949,50,1 \
--distributed-backend nccl \
--lr 0.0001 \
--lr-decay-style linear \
--lr-decay-iters 990000 \
--weight-decay 1e-2 \
--clip-grad 1.0 \
--warmup .01 \
--fp16 \
--fp32-layernorm \
--fp32-embedding \
--skip-mmap-warmup \
--num-workers 0
examples/pretrain_bert.sh
View file @ 5e56e563
...@@ -2,6 +2,8 @@ ...@@ -2,6 +2,8 @@
RANK=0 RANK=0
WORLD_SIZE=1 WORLD_SIZE=1
DATA_PATH=<Specify path and file prefix>_text_sentence
CHECKPOINT_PATH=<Specify path>
python pretrain_bert.py \ python pretrain_bert.py \
--num-layers 24 \ --num-layers 24 \
...@@ -9,26 +11,25 @@ python pretrain_bert.py \ ...@@ -9,26 +11,25 @@ python pretrain_bert.py \
--num-attention-heads 16 \ --num-attention-heads 16 \
--batch-size 4 \ --batch-size 4 \
--seq-length 512 \ --seq-length 512 \
--max-preds-per-seq 80 \
--max-position-embeddings 512 \ --max-position-embeddings 512 \
--train-iters 1000000 \ --train-iters 2000000 \
--save checkpoints/bert_345m \ --save $CHECKPOINT_PATH \
--load checkpoints/bert_345m \ --load $CHECKPOINT_PATH \
--resume-dataloader \ --data-path $DATA_PATH \
--train-data wikipedia \ --vocab-file bert-vocab.txt \
--lazy-loader \ --data-impl mmap \
--tokenizer-type BertWordPieceTokenizer \
--tokenizer-model-type bert-large-uncased \
--presplit-sentences \
--cache-dir cache \
--split 949,50,1 \ --split 949,50,1 \
--distributed-backend nccl \ --distributed-backend nccl \
--lr 0.0001 \ --lr 0.0001 \
--min-lr 0.00001 \
--lr-decay-style linear \ --lr-decay-style linear \
--lr-decay-iters 990000 \ --lr-decay-iters 990000 \
--weight-decay 1e-2 \ --weight-decay 1e-2 \
--clip-grad 1.0 \ --clip-grad 1.0 \
--warmup .01 \ --warmup .01 \
--fp16 \ --log-interval 100 \
--fp32-layernorm \ --save-interval 10000 \
--fp32-embedding --eval-interval 1000 \
--eval-iters 10 \
--fp16
examples/pretrain_bert_distributed.sh
View file @ 5e56e563
...@@ -8,36 +8,37 @@ NNODES=1 ...@@ -8,36 +8,37 @@ NNODES=1
NODE_RANK=0 NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES)) WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES))
DATA_PATH=<Specify path and file prefix>_text_sentence
CHECKPOINT_PATH=<Specify path>
DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT" DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT"
python -m torch.distributed.launch $DISTRIBUTED_ARGS \ python -m torch.distributed.launch $DISTRIBUTED_ARGS \
pretrain_bert.py \ pretrain_bert.py \
--model-parallel-size 1 \
--num-layers 24 \ --num-layers 24 \
--hidden-size 1024 \ --hidden-size 1024 \
--num-attention-heads 16 \ --num-attention-heads 16 \
--batch-size 4 \ --batch-size 4 \
--seq-length 512 \ --seq-length 512 \
--max-preds-per-seq 80 \
--max-position-embeddings 512 \ --max-position-embeddings 512 \
--train-iters 1000000 \ --train-iters 1000000 \
--save checkpoints/bert_345m \ --save $CHECKPOINT_PATH \
--load checkpoints/bert_345m \ --load $CHECKPOINT_PATH \
--resume-dataloader \ --data-path $DATA_PATH \
--train-data wikipedia \ --vocab-file bert-vocab.txt \
--lazy-loader \ --data-impl mmap \
--tokenizer-type BertWordPieceTokenizer \
--tokenizer-model-type bert-large-uncased \
--presplit-sentences \
--cache-dir cache \
--split 949,50,1 \ --split 949,50,1 \
--distributed-backend nccl \ --distributed-backend nccl \
--lr 0.0001 \ --lr 0.0001 \
--lr-decay-style linear \ --lr-decay-style linear \
--min-lr 1.0e-5 \
--lr-decay-iters 990000 \ --lr-decay-iters 990000 \
--weight-decay 1e-2 \ --weight-decay 1e-2 \
--clip-grad 1.0 \ --clip-grad 1.0 \
--warmup .01 \ --warmup .01 \
--fp16 \ --log-interval 100 \
--fp32-layernorm \ --save-interval 10000 \
--fp32-embedding --eval-interval 1000 \
--eval-iters 10 \
--fp16
examples/pretrain_bert_model_parallel.sh deleted 100755 → 0
View file @ 6c0a5bd8
#!/bin/bash
GPUS_PER_NODE=8
# Change for multinode config
MASTER_ADDR=localhost
MASTER_PORT=6000
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES))
DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT"
python -m torch.distributed.launch $DISTRIBUTED_ARGS \
pretrain_bert.py \
--model-parallel-size 2 \
--num-layers 24 \
--hidden-size 1024 \
--num-attention-heads 16 \
--batch-size 4 \
--seq-length 512 \
--max-preds-per-seq 80 \
--max-position-embeddings 512 \
--train-iters 1000000 \
--save checkpoints/bert_345m_mp2 \
--load checkpoints/bert_345m_mp2 \
--resume-dataloader \
--train-data wikipedia \
--lazy-loader \
--tokenizer-type BertWordPieceTokenizer \
--tokenizer-model-type bert-large-uncased \
--presplit-sentences \
--cache-dir cache \
--split 949,50,1 \
--distributed-backend nccl \
--lr 0.0001 \
--lr-decay-style linear \
--lr-decay-iters 990000 \
--weight-decay 1e-2 \
--clip-grad 1.0 \
--warmup .01 \
--fp16 \
--fp32-layernorm \
--fp32-embedding
examples/pretrain_bert_sentencepiece.sh deleted 100755 → 0
View file @ 6c0a5bd8
#!/bin/bash
RANK=0
WORLD_SIZE=1
python pretrain_bert.py \
--num-layers 24 \
--hidden-size 1024 \
--num-attention-heads 16 \
--batch-size 4 \
--seq-length 512 \
--max-preds-per-seq 80 \
--max-position-embeddings 512 \
--train-iters 1000000 \
--save checkpoints/bert_345m \
--load checkpoints/bert_345m \
--resume-dataloader \
--train-data wikipedia \
--lazy-loader \
--tokenizer-type SentencePieceTokenizer \
--tokenizer-model-type bpe \
--tokenizer-path tokenizer.model \
--presplit-sentences \
--cache-dir cache \
--split 949,50,1 \
--distributed-backend nccl \
--lr 0.0001 \
--lr-decay-style linear \
--lr-decay-iters 990000 \
--weight-decay 1e-2 \
--clip-grad 1.0 \
--warmup .01 \
--fp16 \
--fp32-layernorm \
--fp32-embedding
examples/pretrain_bert_tfrecords_distributed.sh deleted 100755 → 0
View file @ 6c0a5bd8
#!/bin/bash
GPUS_PER_NODE=8
# Change for multinode config
MASTER_ADDR=localhost
MASTER_PORT=6000
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES))
DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT"
python -m torch.distributed.launch $DISTRIBUTED_ARGS \
pretrain_bert.py \
--num-layers 24 \
--hidden-size 1024 \
--num-attention-heads 16 \
--batch-size 4 \
--seq-length 512 \
--max-preds-per-seq 80 \
--max-position-embeddings 512 \
--train-iters 1000000 \
--save checkpoints/bert_345m \
--load checkpoints/bert_345m \
--resume-dataloader \
--use-tfrecords \
--train-data <TF Record 1> <TFRecord 2> \
--valid-data <TF Record 3> \
--test-data <TF Record 4> \
--tokenizer-type BertWordPieceTokenizer \
--tokenizer-model-type bert-large-uncased \
--presplit-sentences \
--cache-dir cache \
--split 949,50,1 \
--distributed-backend nccl \
--lr 0.0001 \
--lr-decay-style linear \
--lr-decay-iters 990000 \
--weight-decay 1e-2 \
--clip-grad 1.0 \
--warmup .01 \
--fp16 \
--fp32-layernorm \
--fp32-embedding
examples/pretrain_gpt2.sh
View file @ 5e56e563
...@@ -5,6 +5,10 @@ ...@@ -5,6 +5,10 @@
RANK=0 RANK=0
WORLD_SIZE=1 WORLD_SIZE=1
DATA_PATH=<Specify path and file prefix>_text_document
CHECKPOINT_PATH=<Specify path>
python pretrain_gpt2.py \ python pretrain_gpt2.py \
--num-layers 24 \ --num-layers 24 \
--hidden-size 1024 \ --hidden-size 1024 \
...@@ -12,22 +16,27 @@ python pretrain_gpt2.py \ ...@@ -12,22 +16,27 @@ python pretrain_gpt2.py \
--batch-size 8 \ --batch-size 8 \
--seq-length 1024 \ --seq-length 1024 \
--max-position-embeddings 1024 \ --max-position-embeddings 1024 \
--train-iters 320000 \ --train-iters 500000 \
--save checkpoints/gpt2_345m \ --lr-decay-iters 320000 \
--load checkpoints/gpt2_345m \ --save $CHECKPOINT_PATH \
--resume-dataloader \ --load $CHECKPOINT_PATH \
--train-data wikipedia \ --data-path $DATA_PATH \
--lazy-loader \ --vocab-file gpt2-vocab.json \
--tokenizer-type GPT2BPETokenizer \ --merge-file gpt2-merges.txt \
--cache-dir cache \ --data-impl mmap \
--split 949,50,1 \ --split 949,50,1 \
--distributed-backend nccl \ --distributed-backend nccl \
--lr 0.00015 \ --lr 0.00015 \
--min-lr 1.0e-5 \
--lr-decay-style cosine \ --lr-decay-style cosine \
--weight-decay 1e-2 \ --weight-decay 1e-2 \
--clip-grad 1.0 \ --clip-grad 1.0 \
--warmup .01 \ --warmup .01 \
--checkpoint-activations \ --checkpoint-activations \
--log-interval 100 \
--save-interval 10000 \
--eval-interval 1000 \
--eval-iters 10 \
--fp16 --fp16
......
examples/pretrain_gpt2_distributed.sh
View file @ 5e56e563
...@@ -10,33 +10,43 @@ NNODES=1 ...@@ -10,33 +10,43 @@ NNODES=1
NODE_RANK=0 NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES)) WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES))
DATA_PATH=<Specify path and file prefix>_text_document
CHECKPOINT_PATH=<Specify path>
DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT" DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT"
python -m torch.distributed.launch $DISTRIBUTED_ARGS \ python -m torch.distributed.launch $DISTRIBUTED_ARGS \
pretrain_gpt2.py \ pretrain_gpt2.py \
--model-parallel-size 1 \
--num-layers 24 \ --num-layers 24 \
--hidden-size 1024 \ --hidden-size 1024 \
--num-attention-heads 16 \ --num-attention-heads 16 \
--batch-size 8 \ --batch-size 8 \
--seq-length 1024 \ --seq-length 1024 \
--max-position-embeddings 1024 \ --max-position-embeddings 1024 \
--train-iters 320000 \ --train-iters 500000 \
--save checkpoints/gpt2_345m \ --lr-decay-iters 320000 \
--load checkpoints/gpt2_345m \ --save $CHECKPOINT_PATH \
--resume-dataloader \ --load $CHECKPOINT_PATH \
--train-data wikipedia \ --data-path $DATA_PATH \
--lazy-loader \ --vocab-file gpt2-vocab.json \
--tokenizer-type GPT2BPETokenizer \ --merge-file gpt2-merges.txt \
--cache-dir cache \ --data-impl mmap \
--split 949,50,1 \ --split 949,50,1 \
--distributed-backend nccl \ --distributed-backend nccl \
--lr 0.00015 \ --lr 0.00015 \
--lr-decay-style cosine \ --lr-decay-style cosine \
--min-lr 1.0e-5 \
--weight-decay 1e-2 \ --weight-decay 1e-2 \
--clip-grad 1.0 \ --clip-grad 1.0 \
--warmup .01 \ --warmup .01 \
--checkpoint-activations \ --checkpoint-activations \
--log-interval 100 \
--save-interval 10000 \
--eval-interval 1000 \
--eval-iters 10 \
--fp16 --fp16
set +x set +x
examples/pretrain_gpt2_model_parallel.sh deleted 100755 → 0
View file @ 6c0a5bd8
#! /bin/bash
# Runs the "345M" parameter model
GPUS_PER_NODE=8
# Change for multinode config
MASTER_ADDR=localhost
MASTER_PORT=6000
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES))
DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT"
python -m torch.distributed.launch $DISTRIBUTED_ARGS \
pretrain_gpt2.py \
--model-parallel-size 2 \
--num-layers 24 \
--hidden-size 1024 \
--num-attention-heads 16 \
--batch-size 8 \
--seq-length 1024 \
--max-position-embeddings 1024 \
--train-iters 320000 \
--save checkpoints/gpt2_345m_mp2 \
--load checkpoints/gpt2_345m_mp2 \
--resume-dataloader \
--train-data wikipedia \
--lazy-loader \
--tokenizer-type GPT2BPETokenizer \
--cache-dir cache \
--split 949,50,1 \
--distributed-backend nccl \
--lr 0.00015 \
--lr-decay-style cosine \
--weight-decay 1e-2 \
--clip-grad 1.0 \
--warmup .01 \
--checkpoint-activations \
--fp16
set +x
examples/run_gpt2_eval.py deleted 100644 → 0
View file @ 6c0a5bd8
"""
example usage:
python scripts/run_gpt2_eval.py \
--model-parallel-size 1 \
--num-layers 12 \
--hidden-size 768 \
--num-attention-heads 12 \
--model-path <gpt2_117_path> \
--data-path <wikitext_tokens_test_path> \
--batch-size 16 \
--cache-dir <cache dir path>
"""
import argparse
import subprocess
parser = argparse.ArgumentParser('run zero shot GPT2 eval')
parser.add_argument('--model-path', type=str, required=True,
help='Saved model path for evaluation')
parser.add_argument('--batch-size', type=int, default=4,
help='batch size to use for evaluation')
parser.add_argument('--num-attention-heads', type=int, default=12,
help='num of transformer attention heads')
parser.add_argument('--hidden-size', type=int, default=768,
help='tansformer hidden size')
parser.add_argument('--num-layers', type=int, default=12,
help='num decoder layers')
parser.add_argument('--data-path', type=str, required=True,
help='Data path for evaluation data')
parser.add_argument('--cloze-eval', action='store_true',
help='Run lambada cloze eval instead of perplexity eval.')
parser.add_argument('--easy-lambada', action='store_true',
help='use easier formulation of lambada')
parser.add_argument('--model-parallel-size', type=int, default=1,
help='model parallel size to use')
args = parser.parse_args()
multinode_args = ''
if args.model_parallel_size > 1:
multinode_args += ' -m torch.distributed.launch --nproc_per_node {} '.format(args.model_parallel_size)
CMD = ' --model-parallel-size {model_par} \
--num-layers {nlayers} \
--hidden-size {hidden} \
--log-interval 100 \
--load {model} \
--batch-size {batch} \
--num-attention-heads {natt} \
--seq-length 1024 \
--max-position-embeddings 1024 \
--tokenizer-type GPT2BPETokenizer \
--distributed-backend nccl \
--hidden-dropout 0.1 \
--attention-dropout 0.1 \
--fp16 \
--lr 1 --no-load-optim --no-load-rng --epochs 0 \
--overlapping-eval 32 \
--merge-file /home/universal-lm-data.cosmos549/repos/megatron_latest/vocab_cache/merges.txt \
--vocab-file /home/universal-lm-data.cosmos549/repos/megatron_latest/vocab_cache/vocab.json'.format(model_par=args.model_parallel_size,
nlayers=args.num_layers,
hidden=args.hidden_size,
model=args.model_path,
batch=args.batch_size,
natt=args.num_attention_heads,)
if args.cloze_eval:
CMD += ' --valid-data {} '.format(args.data_path)
CMD += ' --task LAMBADA '
if not args.easy_lambada:
CMD += ' --strict-lambada '
CMD = 'main.py' + CMD
print('Running Lambada Eval Command:', flush=True)
else:
CMD += ' --valid-data {} '.format(args.data_path)
CMD += ' --task WIKITEXT103 '
CMD = 'main.py' + CMD
print('Running PPL Eval Command:', flush=True)
CMD = 'python3 '+multinode_args+CMD
print(CMD, flush=True)
subprocess.call(CMD.split())
megatron/__init__.py
View file @ 5e56e563
# coding=utf-8 # coding=utf-8
# Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved. # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
# #
# Licensed under the Apache License, Version 2.0 (the "License"); # Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License. # you may not use this file except in compliance with the License.
......
megatron/arguments.py
View file @ 5e56e563
# coding=utf-8 # coding=utf-8
# Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved. # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
# #
# Licensed under the Apache License, Version 2.0 (the "License"); # Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License. # you may not use this file except in compliance with the License.
...@@ -19,9 +19,11 @@ import argparse ...@@ -19,9 +19,11 @@ import argparse
import os import os
def parse_args(extra_args_provider=None, defaults={}): def parse_args(extra_args_provider=None, defaults={},
ignore_unknown_args=False):
"""Parse all arguments.""" """Parse all arguments."""
parser = argparse.ArgumentParser(description='Megatron-LM Arguments') parser = argparse.ArgumentParser(description='Megatron-LM Arguments',
allow_abbrev=False)
# Standard arguments. # Standard arguments.
parser = _add_network_size_args(parser) parser = _add_network_size_args(parser)
...@@ -35,24 +37,16 @@ def parse_args(extra_args_provider=None, defaults={}): ...@@ -35,24 +37,16 @@ def parse_args(extra_args_provider=None, defaults={}):
parser = _add_validation_args(parser) parser = _add_validation_args(parser)
parser = _add_data_args(parser) parser = _add_data_args(parser)
parser = _add_autoresume_args(parser) parser = _add_autoresume_args(parser)
# TODO: Refactor
parser = _add_gpt2_args(parser)
# Custom arguments. # Custom arguments.
if extra_args_provider is not None: if extra_args_provider is not None:
parser = extra_args_provider(parser) parser = extra_args_provider(parser)
# Parse. # Parse.
args = parser.parse_args() if ignore_unknown_args:
args, _ = parser.parse_known_args()
# Set input defaults. else:
for key in defaults: args = parser.parse_args()
# For default to be valid, it should not be provided in the
# arguments that are passed to the program. We check this by
# ensuring the arg is set to None.
assert getattr(args, key) is None, \
'defaults can only be overwritten for args with None values.'
setattr(args, key, defaults[key])
# Distributed args. # Distributed args.
args.rank = int(os.getenv('RANK', '0')) args.rank = int(os.getenv('RANK', '0'))
...@@ -67,6 +61,26 @@ def parse_args(extra_args_provider=None, defaults={}): ...@@ -67,6 +61,26 @@ def parse_args(extra_args_provider=None, defaults={}):
if args.loss_scale is None: if args.loss_scale is None:
args.dynamic_loss_scale = True args.dynamic_loss_scale = True
# Set input defaults.
for key in defaults:
# For default to be valid, it should not be provided in the
# arguments that are passed to the program. We check this by
# ensuring the arg is set to None.
if getattr(args, key) is not None:
if args.rank == 0:
print('WARNING: overriding default arguments for {key}:{v} \
with {key}:{v2}'.format(key=key, v=defaults[key],
v2=getattr(args, key)),
flush=True)
else:
setattr(args, key, defaults[key])
# Check required arguments.
required_args = ['num_layers', 'hidden_size', 'num_attention_heads',
'max_position_embeddings']
for req_arg in required_args:
_check_arg_is_not_none(args, req_arg)
# Checks. # Checks.
assert args.hidden_size % args.num_attention_heads == 0 assert args.hidden_size % args.num_attention_heads == 0
if args.seq_length is not None: if args.seq_length is not None:
...@@ -93,16 +107,20 @@ def _print_args(args): ...@@ -93,16 +107,20 @@ def _print_args(args):
print('---------------- end of arguments ----------------', flush=True) print('---------------- end of arguments ----------------', flush=True)
def _check_arg_is_not_none(args, arg):
assert getattr(args, arg) is not None, '{} argument is None'.format(arg)
def _add_network_size_args(parser): def _add_network_size_args(parser):
group = parser.add_argument_group(title='network size') group = parser.add_argument_group(title='network size')
group.add_argument('--num-layers', type=int, required=True, group.add_argument('--num-layers', type=int, default=None,
help='Number of transformer layers.') help='Number of transformer layers.')
group.add_argument('--hidden-size', type=int, required=True, group.add_argument('--hidden-size', type=int, default=None,
help='Tansformer hidden size.') help='Tansformer hidden size.')
group.add_argument('--num-attention-heads', type=int, required=True, group.add_argument('--num-attention-heads', type=int, default=None,
help='Number of transformer attention heads.') help='Number of transformer attention heads.')
group.add_argument('--max-position-embeddings', type=int, required=True, group.add_argument('--max-position-embeddings', type=int, default=None,
help='Maximum number of position embeddings to use. ' help='Maximum number of position embeddings to use. '
'This is the size of position embedding.') 'This is the size of position embedding.')
group.add_argument('--make-vocab-size-divisible-by', type=int, default=128, group.add_argument('--make-vocab-size-divisible-by', type=int, default=128,
...@@ -114,6 +132,10 @@ def _add_network_size_args(parser): ...@@ -114,6 +132,10 @@ def _add_network_size_args(parser):
action='store_true', action='store_true',
help='If set, use original BERT residula connection ' help='If set, use original BERT residula connection '
'ordering.') 'ordering.')
group.add_argument('--openai-gelu', action='store_true',
help='Use OpenAIs GeLU implementation. This option'
'should not be used unless for backward compatibility'
'reasons.')
return parser return parser
...@@ -322,6 +344,7 @@ def _add_data_args(parser): ...@@ -322,6 +344,7 @@ def _add_data_args(parser):
group.add_argument('--tokenizer-type', type=str, group.add_argument('--tokenizer-type', type=str,
default=None, default=None,
choices=['BertWordPieceLowerCase', choices=['BertWordPieceLowerCase',
'BertWordPieceCase',
'GPT2BPETokenizer'], 'GPT2BPETokenizer'],
help='What type of tokenizer to use.') help='What type of tokenizer to use.')
group.add_argument('--data-impl', type=str, default='infer', group.add_argument('--data-impl', type=str, default='infer',
...@@ -348,31 +371,3 @@ def _add_autoresume_args(parser): ...@@ -348,31 +371,3 @@ def _add_autoresume_args(parser):
'termination signal') 'termination signal')
return parser return parser
########################################################################
def _add_gpt2_args(parser):
group = parser.add_argument_group(title='gpt2')
group.add_argument('--input-data-sizes-file', type=str, default='sizes.txt',
help='The filename containing all the shards '
'sizes for numpy data loader')
return parser
def add_data_args_(parser):
"""Train/valid/test data arguments."""
group = parser.add_argument_group('data', 'data configurations')
group.add_argument('--data-loader', type=str, default=None,
choices=['raw', 'lazy', 'tfrecords', 'numpy', 'binary'],
help='Which data loader to use. Default varies by model.')
return parser
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment