Skip to content

GitLab

Commit 050c24d5 authored by Steven Steinke's avatar Steven Steinke
Browse files

Merge branch 'staging_docs' into 'staging' 

Staging docs

See merge request ADLR/megatron-lm!64
parents d928dd51 3df23c72
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 667 additions and 493 deletions
README.md
View file @ 050c24d5
This diff is collapsed.
examples/evaluate_zeroshot_gpt2.sh 0 → 100755
View file @ 050c24d5
#!/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 512 \
--max-position-embeddings 512 \
--log-interval 10 \
--fp16 \
--no-load-optim \
--no-load-rng
examples/finetine_race_distributed.sh 0 → 100755
View file @ 050c24d5
#!/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
PRETRIANED_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 $PRETRIANED_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 500000 \
--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/finetune_mnli_distributed.sh 0 → 100755
View file @ 050c24d5
#!/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/generate_text.sh
View file @ 050c24d5
#!/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 $TOPP \ --genfile unconditional_samples.json \
--num-samples 2 \
--top_p 0.9 \
--recompute --recompute
examples/merge_mp_bert.sh 0 → 100755
View file @ 050c24d5
#!/bin/bash
MODEL_PARALLEL_SIZE=2
VOCAB_FILE=bert-vocab.txt
CHECKPOINT_PATH=checkpoints/bert_345m
WORLD_SIZE=$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 $CHECKPOINT_PATH
examples/pretrain_albert.sh deleted 100755 → 0
View file @ d928dd51
#!/bin/bash
RANK=0
WORLD_SIZE=1
python pretrain_albert.py \
--num-layers 12 \
--hidden-size 768 \
--num-attention-heads 12 \
--batch-size 4 \
--seq-length 512 \
--max-preds-per-seq 80 \
--max-position-embeddings 512 \
--train-iters 10000 \
--save checkpoints/albert_117m \
--load checkpoints/albert_117m \
--resume-dataloader \
--data-path data/megatron/bc_rn_owt_sto_wiki_dedup_shuf_cleaned_0.7_mmap \
--vocab data/megatron/vocab.txt \
--split 949,50,1 \
--distributed-backend nccl \
--lr 0.0001 \
--lr-decay-style linear \
--lr-decay-iters 990000 \
--weight-decay 1e-2 \
--clip-grad 1.0 \
--warmup .01 \
--fp16 \
--fp32-layernorm \
--fp32-embedding \
--skip-mmap-warmup \
--num-workers 0
examples/pretrain_albert_distributed.sh deleted 100755 → 0
View file @ d928dd51
#!/bin/bash
GPUS_PER_NODE=2
# Change for multinode config
MASTER_ADDR=localhost
MASTER_PORT=6000
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$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 @ 050c24d5
...@@ -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 $DATA_PATH \
--train-data wikipedia \ --vocab-file bert-vocab.txt \
--lazy-loader \ --data-impl mmap \
--tokenizer-type BertWordPieceTokenizer \
--tokenizer-model-type bert-large-uncased \
--presplit-sentences \
--cache-dir cache \
--split 949,50,1 \ --split 949,50,1 \
--distributed-backend nccl \ --distributed-backend nccl \
--lr 0.0001 \ --lr 0.0001 \
--min-lr 0.00001 \
--lr-decay-style linear \ --lr-decay-style linear \
--lr-decay-iters 990000 \ --lr-decay-iters 990000 \
--weight-decay 1e-2 \ --weight-decay 1e-2 \
--clip-grad 1.0 \ --clip-grad 1.0 \
--warmup .01 \ --warmup .01 \
--fp16 \ --log-interval 100 \
--fp32-layernorm \ --save-interval 10000 \
--fp32-embedding --eval-interval 1000 \
--eval-iters 10 \
--fp16
examples/pretrain_bert_distributed.sh
View file @ 050c24d5
...@@ -8,36 +8,37 @@ NNODES=1 ...@@ -8,36 +8,37 @@ NNODES=1
NODE_RANK=0 NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$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 $DATA_PATH \
--train-data wikipedia \ --vocab-file bert-vocab.txt \
--lazy-loader \ --data-impl mmap \
--tokenizer-type BertWordPieceTokenizer \
--tokenizer-model-type bert-large-uncased \
--presplit-sentences \
--cache-dir cache \
--split 949,50,1 \ --split 949,50,1 \
--distributed-backend nccl \ --distributed-backend nccl \
--lr 0.0001 \ --lr 0.0001 \
--lr-decay-style linear \ --lr-decay-style linear \
--min-lr 1.0e-5 \
--lr-decay-iters 990000 \ --lr-decay-iters 990000 \
--weight-decay 1e-2 \ --weight-decay 1e-2 \
--clip-grad 1.0 \ --clip-grad 1.0 \
--warmup .01 \ --warmup .01 \
--fp16 \ --log-interval 100 \
--fp32-layernorm \ --save-interval 10000 \
--fp32-embedding --eval-interval 1000 \
--eval-iters 10 \
--fp16
examples/pretrain_bert_model_parallel.sh deleted 100755 → 0
View file @ d928dd51
#!/bin/bash
GPUS_PER_NODE=8
# Change for multinode config
MASTER_ADDR=localhost
MASTER_PORT=6000
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$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 $DISTRIBUTED_ARGS \
pretrain_bert.py \
--model-parallel-size 2 \
--num-layers 24 \
--hidden-size 1024 \
--num-attention-heads 16 \
--batch-size 4 \
--seq-length 512 \
--max-preds-per-seq 80 \
--max-position-embeddings 512 \
--train-iters 1000000 \
--save checkpoints/bert_345m_mp2 \
--load checkpoints/bert_345m_mp2 \
--resume-dataloader \
--train-data wikipedia \
--lazy-loader \
--tokenizer-type BertWordPieceTokenizer \
--tokenizer-model-type bert-large-uncased \
--presplit-sentences \
--cache-dir cache \
--split 949,50,1 \
--distributed-backend nccl \
--lr 0.0001 \
--lr-decay-style linear \
--lr-decay-iters 990000 \
--weight-decay 1e-2 \
--clip-grad 1.0 \
--warmup .01 \
--fp16 \
--fp32-layernorm \
--fp32-embedding
examples/pretrain_bert_sentencepiece.sh deleted 100755 → 0
View file @ d928dd51
#!/bin/bash
RANK=0
WORLD_SIZE=1
python pretrain_bert.py \
--num-layers 24 \
--hidden-size 1024 \
--num-attention-heads 16 \
--batch-size 4 \
--seq-length 512 \
--max-preds-per-seq 80 \
--max-position-embeddings 512 \
--train-iters 1000000 \
--save checkpoints/bert_345m \
--load checkpoints/bert_345m \
--resume-dataloader \
--train-data wikipedia \
--lazy-loader \
--tokenizer-type SentencePieceTokenizer \
--tokenizer-model-type bpe \
--tokenizer-path tokenizer.model \
--presplit-sentences \
--cache-dir cache \
--split 949,50,1 \
--distributed-backend nccl \
--lr 0.0001 \
--lr-decay-style linear \
--lr-decay-iters 990000 \
--weight-decay 1e-2 \
--clip-grad 1.0 \
--warmup .01 \
--fp16 \
--fp32-layernorm \
--fp32-embedding
examples/pretrain_bert_tfrecords_distributed.sh deleted 100755 → 0
View file @ d928dd51
#!/bin/bash
GPUS_PER_NODE=8
# Change for multinode config
MASTER_ADDR=localhost
MASTER_PORT=6000
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$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 @ 050c24d5
...@@ -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 $DATA_PATH \
--lazy-loader \ --vocab-file gpt2-vocab.json \
--tokenizer-type GPT2BPETokenizer \ --merge-file gpt2-merges.txt \
--cache-dir cache \ --data-impl mmap \
--split 949,50,1 \ --split 949,50,1 \
--distributed-backend nccl \ --distributed-backend nccl \
--lr 0.00015 \ --lr 0.00015 \
--min-lr 1.0e-5 \
--lr-decay-style cosine \ --lr-decay-style cosine \
--weight-decay 1e-2 \ --weight-decay 1e-2 \
--clip-grad 1.0 \ --clip-grad 1.0 \
--warmup .01 \ --warmup .01 \
--checkpoint-activations \ --checkpoint-activations \
--log-interval 100 \
--save-interval 10000 \
--eval-interval 1000 \
--eval-iters 10 \
--fp16 --fp16
......
examples/pretrain_gpt2_distributed.sh
View file @ 050c24d5
...@@ -10,33 +10,43 @@ NNODES=1 ...@@ -10,33 +10,43 @@ NNODES=1
NODE_RANK=0 NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$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 $DATA_PATH \
--lazy-loader \ --vocab-file gpt2-vocab.json \
--tokenizer-type GPT2BPETokenizer \ --merge-file gpt2-merges.txt \
--cache-dir cache \ --data-impl mmap \
--split 949,50,1 \ --split 949,50,1 \
--distributed-backend nccl \ --distributed-backend nccl \
--lr 0.00015 \ --lr 0.00015 \
--lr-decay-style cosine \ --lr-decay-style cosine \
--min-lr 1.0e-5 \
--weight-decay 1e-2 \ --weight-decay 1e-2 \
--clip-grad 1.0 \ --clip-grad 1.0 \
--warmup .01 \ --warmup .01 \
--checkpoint-activations \ --checkpoint-activations \
--log-interval 100 \
--save-interval 10000 \
--eval-interval 1000 \
--eval-iters 10 \
--fp16 --fp16
set +x set +x
examples/pretrain_gpt2_model_parallel.sh deleted 100755 → 0
View file @ d928dd51
#! /bin/bash
# Runs the "345M" parameter model
GPUS_PER_NODE=8
# Change for multinode config
MASTER_ADDR=localhost
MASTER_PORT=6000
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$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
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