split_tp_partitions_102B.sh 2 KB
Newer Older
Rayyyyy's avatar
Rayyyyy committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
#!/bin/bash

#merge checkpoint along the pipeline

LOAD_CHECKPOINT_PATH=$1
#<Specify the loaded ckpt path>
SAVE_CHECKPOINT_PATH=$2
#<Specify the stored ckpt path>
TOKENIZER_MODEL_PATH=$3
#<Specify tokenizer model path>
export CUDA_DEVICE_MAX_CONNECTIONS=1
export PATH=/opt/conda/bin/:$PATH

if [ ! -d $SAVE_CHECKPOINT_PATH ]; then

        mkdir $SAVE_CHECKPOINT_PATH

fi

python tools/split_tp_partitions.py \
    --tokenizer-model-path $TOKENIZER_MODEL_PATH \
    --tensor-model-parallel-size 1 \
    --target-tensor-model-parallel-size 8 \
    --pipeline-model-parallel-size 32 \
    --target-pipeline-model-parallel-size 32 \
    --pipeline-model-parallel-method block \
    --pipeline-model-parallel-blocks 2,2,2,2,2,2,2,2,2,2,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,2,2 \
    --pipeline-generate-layer 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31 \
    --tokenizer-type YuanTokenizer \
    --num-layers 84 \
    --hidden-size 8192 \
    --num-attention-heads 64 \
    --seq-length 4096 \
    --max-position-embeddings 4096 \
    --use-lf-gate \
    --lf-conv2d-group 1 \
    --lf-conv2d-num-pad 1 \
    --position-embedding-type rope \
    --flash-attn-drop 0.1\
    --fim-rate 0.5\
    --fim-spm-rate 0.5\
    --attention-dropout 0\
    --hidden-dropout 0\
    --norm-dtype RMSNorm \
    --disable-bias-linear \
    --reset-position-ids \
    --use-flash-attn \
    --swiglu \
    --fused-rmsnorm \
    --DDP-impl local \
    --bf16 \
    --process-checkpoint \
    --save-interval 1 \
    --recompute-method block \
    --recompute-granularity full \
    --recompute-num-layers 2 \
    --load $LOAD_CHECKPOINT_PATH \
    --save $SAVE_CHECKPOINT_PATH \
    --micro-batch-size 1 \
    --global-batch-size 1152 \
    --no-load-optim \
    --use-distributed-optimizer \
    --lr 0.0001 \
    --train-iters 63578 \
    --lr-decay-iters 63578 \
    --lr-decay-style cosine \
    --min-lr 1.0e-5 \
    --weight-decay 1e-1 \
    --use-cpu-initialization \
    --data-impl mmap
du -sh $SAVE_CHECKPOINT_PATH