pretrain_gpt.sh 1.04 KB
Newer Older
1
2
3
4
5
6
7
#! /bin/bash

# Runs the "345M" parameter model

RANK=0
WORLD_SIZE=1

Mohammad's avatar
Mohammad committed
8
9
10
11
DATA_PATH=<Specify path and file prefix>_text_document
CHECKPOINT_PATH=<Specify path>


12
python pretrain_gpt.py \
13
14
15
       --num-layers 24 \
       --hidden-size 1024 \
       --num-attention-heads 16 \
16
17
       --micro-batch-size 4 \
       --global-batch-size 8 \
18
19
       --seq-length 1024 \
       --max-position-embeddings 1024 \
Mohammad's avatar
Mohammad committed
20
21
22
23
24
25
26
27
       --train-iters 500000 \
       --lr-decay-iters 320000 \
       --save $CHECKPOINT_PATH \
       --load $CHECKPOINT_PATH \
       --data-path $DATA_PATH \
       --vocab-file gpt2-vocab.json \
       --merge-file gpt2-merges.txt \
       --data-impl mmap \
28
29
30
       --split 949,50,1 \
       --distributed-backend nccl \
       --lr 0.00015 \
Mohammad's avatar
Mohammad committed
31
       --min-lr 1.0e-5 \
32
33
34
       --lr-decay-style cosine \
       --weight-decay 1e-2 \
       --clip-grad 1.0 \
35
       --lr-warmup-fraction .01 \
36
       --activations-checkpoint-method uniform \
Mohammad's avatar
Mohammad committed
37
38
39
40
       --log-interval 100 \
       --save-interval 10000 \
       --eval-interval 1000 \
       --eval-iters 10 \
41
       --fp16