deep_search/sft/mix_math_multi_node.sh

#!/bin/bash

export OMP_NUM_THREADS=20
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
# 定义参数
lr=1e-5
base=QwQ-32B
tokenizer=QwQ-32B
# train_data=hopotqa_1217.json
train_data=no_error_data_871
bsz=1
acc=8

# 生成随机 JOB-ID
JOB_ID=$(( RANDOM % 100000 )) # 生成一个 0 到 99999 的随机数
save_path="JOB:${JOB_ID}#LR:${lr}#BASE:${base}#TOKEN:${tokenizer}#BSZ:${bsz}#ACC:${acc}_${train_data}_mixed_math"

# 输出路径
output_dir="/capacity/userdata/ss/sft_search/${save_path}"

output_dir_1=${output_dir}
model_name_1=${base}
dataset_1=${train_data}
# 创建输出目录
mkdir -p "$output_dir"

echo ${output_dir}

# 执行 deepspeed 命令
/opt/aps/workdir/miniforge3/envs/ss_train/bin/deepspeed \
    --hostfile=hostfile \
    --no_ssh \
    --node_rank=0 \
    --master_addr=172.19.164.116 \
    --master_port=9944 \
    sft_2_math.py \
    --deepspeed ds_zero3_offload.json \
    --model_name_or_path "/capacity/userdata/models/${base}" \
    --tokenizer_name_or_path "/capacity/userdata/models/${tokenizer}" \
    --do_train \
    --save_safetensors true \
    --data_path "/opt/aps/workdir/sunshuang/deep_search/search_o1/sft_data/${train_data}.json" \
    --lr_scheduler_type cosine \
    --output_dir "$output_dir" \
    --overwrite_output_dir \
    --warmup_ratio 0.03 \
    --gradient_checkpointing true \
    --per_device_train_batch_size "$bsz" \
    --gradient_accumulation_steps "$acc" \
    --logging_steps 1 \
    --learning_rate "$lr" \
    --num_train_epochs 6 \
    --save_strategy epoch \
    --save_only_model true \
    --model_max_length 30000 \
    --save_total_limit 6 \
    --bf16 || exit 1


bash test_two_model_qwq.sh $output_dir_1 $model_name_1 $dataset_1
# bash test.sh $output_dir_2 $model_name_2
# bash test.sh $output_dir_3 $model_name_3

# bash test.sh $output_dir_2 $model_name_2