pretrain_gpt2_model_parallel.sh

#! /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