multi_dcu_train.sh

#!/bin/bash

deepspeed --master_port $(shuf -n 1 -i 10000-65535) --include="localhost:3,4" /DISC-FinLLM_ModelZoo/LLaMA-Factory/src/train_bash.py \
    --deepspeed /DISC-FinLLM_ModelZoo/LLaMA-Factory/deepspeed.json \
    --stage sft \
    --do_train \
    --model_name_or_path /DISC-FinLLM_ModelZoo/FinLLM \
    --dataset retrieval_part \
    --dataset_dir /DISC-FinLLM_ModelZoo/LLaMA-Factory/data \
    --template baichuan2 \
    --finetuning_type lora \
    --lora_target all \
    --output_dir /DISC-FinLLM_ModelZoo/saves/lora_multi \
    --overwrite_output_dir \
    --cutoff_len 1024 \
    --preprocessing_num_workers 4 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --warmup_steps 20 \
    --save_steps 100 \
    --eval_steps 100 \
    --evaluation_strategy steps \
    --learning_rate 5e-5 \
    --num_train_epochs 10.0 \
    --max_samples 3000 \
    --val_size 0.1 \
    --ddp_timeout 180000000 \
    --plot_loss \
    --fp16