# 程式碼-多顆GPUs微調模型

## 多顆GPUs微調模型

## :memo: Where do I start?
- 聯絡窗口 Email us : [e-mail] 王小姐

### 事前準備
- 映像檔下載

singularity pull c00cjz00/c00cjz00_cuda11.8_pytorch:2.1.2-cuda11.8-cudnn8-devel-llama_factory

- 原生安裝所需套件

# 套件安裝ubuntu相關套件
apt install libfontconfig  libaio-dev  libibverbs-dev jq

# 安裝 Llama-factory 相關套件
pip install llmtuner==0.5.3 deepspeed==0.13.1  bitsandbytes==0.42.0 opencc opencc-python-reimplemented

### 程式碼 C02_LLaMA-Factory-Cmd-lora_nGPUs.ipynb

%%bash
## 請記得修HF_TOKEN='hf_' , 給予huggingface token
## 將模型微調指令寫成 demo.cmd
## 特別注意虛擬絕對路徑部分 /DEEPSPEED/LLaMA-Factory (為你目前執行目錄下必須有LLaMA-Factory目錄)
## --model_name_or_path meta-llama/Llama-2-7b-hf
## --model_name_or_path /work/u00cjz00/slurm_jobs/github/models/Llama-2-7b-chat-hf
##     --deepspeed examples/full_multi_gpu/ds_z3_config.json \

cat << \EOF >  demo.cmd
#!/bin/bash

# 虛擬絕對路徑
cd /DEEPSPEED/LLaMA-Factory

## 清空之前計算
ps -ef |grep 'train_bash.py' | awk '{print $2}' | xargs kill -9
rm -rf ../saves/LLaMA2-7B/lora/sft
sleep 10

## 微調程式
deepspeed --num_gpus ${GPUS_PER_NODE} src/train_bash.py \
    --deepspeed examples/full_multi_gpu/ds_z2_config.json \
    --stage sft \
    --do_train \
    --model_name_or_path /work/u00cjz00/slurm_jobs/github/models/Llama-2-7b-chat-hf \
    --dataset alpaca_gpt4_zh \
    --dataset_dir data \
    --template default \
    --finetuning_type lora \
    --lora_target q_proj,v_proj \
    --output_dir ../saves/LLaMA2-7B/lora/sft \
    --overwrite_cache \
    --overwrite_output_dir \
    --cutoff_len 1024 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 8 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 100 \
    --eval_steps 100 \
    --evaluation_strategy steps \
    --load_best_model_at_end \
    --learning_rate 5e-5 \
    --num_train_epochs 3.0 \
    --max_samples 1000 \
    --val_size 0.1 \
    --plot_loss \
    --fp16

EOF

## 改變執行檔權限並印出
chmod 755 demo.cmd
#cat demo.cmd

- slrum job

%%bash
# SLURM 工作配送 
# --nodes=2 (2台電腦) 
# --gres=gpu:2 (每台電腦2顆GPU)
# -c 8  (每台電腦　8 CORE CPU)

## 將 << \EOF > .... EOF  之間文字 儲存成 demo.slurm
cat << \EOF >  demo.slurm
#!/bin/bash
#SBATCH -A MST110386                                                    ### project number, Example MST109178
#SBATCH -J _t2demo_                                                     ### Job name, Exmaple jupyterlab
#SBATCH -p gp4d                                                         ### Partition Name, Example ngs1gpu
#SBATCH --nodes=1                                                       ### Nodes, Default 1, node number
#SBATCH --ntasks-per-node=1                                             ### Tasks, Default 1, per node tasks
#SBATCH -c 32                                                            ### Cores assigned to each task, Example 4
#SBATCH --gres=gpu:8                                                    ### GPU number, Example gpu:1
#SBATCH --time=0-1:00:00                                                ### Runnung time, days-hours:minutes:seconds or hours:minutes:seconds
#SBATCH -o demo.out                                                     ### Log folder, Here %j is job ID
#SBATCH -e demo.err 

## 環境變數
export GPUS_PER_NODE=8

## 用Singularity 容器執行上方demo.cmd
## $PWD:/DEEPSPEED 交目前目錄掛載成虛擬目錄 /DEEPSPEED, 對應上個CELL  虛擬絕對路徑 cd /DEEPSPEED/LLaMA-Factory
srun singularity exec --nv \
-B $PWD:/DEEPSPEED \
-B /work/u00cjz00/slurm_jobs/github/models/Llama-2-7b-chat-hf \
/work/u00cjz00/nvidia/cuda118/c00cjz00_cuda11.8_pytorch_2.1.2-cuda11.8-cudnn8-devel-llama_factory.sif \
bash -c '/DEEPSPEED/demo.cmd'
EOF

## 改變執行檔權限並印出
chmod 755 demo.slurm
#cat demo.slurm

- 開始配送工作 (請將下面 # 取消, 進行工作派送)
執行後會印出 JOBID, 可於中途刪除工作時使用

!sbatch  demo.slurm

- 觀看派送結果

!echo "觀看派送結果"
!squeue -u $(whoami)

- 印出運算過程訊息

!echo "印出派送運算訊息"
!tail -f demo.out demo.err

- 刪除工作 scancel $JOBID

!scancel $JOBID                              

### dz_z2.config.json

{
  "train_batch_size": "auto",
  "train_micro_batch_size_per_gpu": "auto",
  "gradient_accumulation_steps": "auto",
  "gradient_clipping": "auto",
  "zero_allow_untested_optimizer": true,
  "fp16": {
    "enabled": "auto",
    "loss_scale": 0,
    "loss_scale_window": 1000,
    "initial_scale_power": 16,
    "hysteresis": 2,
    "min_loss_scale": 1
  },
  "bf16": {
    "enabled": "auto"
  },
  "zero_optimization": {
    "stage": 2,
    "allgather_partitions": true,
    "allgather_bucket_size": 5e8,
    "overlap_comm": true,
    "reduce_scatter": true,
    "reduce_bucket_size": 5e8,
    "contiguous_gradients": true,
    "round_robin_gradients": true
  }
}