jd_train/mm-zero3.sh

25 lines
946 B
Bash
Executable File

# 统一环境(会被 deepspeed 的 ssh 继承到各节点)
unset DS_BUILD_OPS DS_SKIP_CUDA_BUILD PYTHONNOUSERSITE
export TORCH_EXTENSIONS_DIR=/tmp/$USER/torch_ext
export PYTORCH_CUDA_ALLOC_CONF="max_split_size_mb:128,expandable_segments:True,garbage_collection_threshold:0.9"
deepspeed --hostfile hostfile \
--num_nodes 6 --num_gpus 4 \
/home/test/jd_train/train_sft_ds.py \
--model_name_or_path /home/test/Qwen3-32B \
--data_glob "/home/test/datasets/my_corpus/train.jsonl" \
--output_dir /home/test/checkpoints/q3-32b-ds4 \
--seq_len 512 \
--per_device_train_batch_size 1 \
--gradient_accumulation_steps 1 \
--learning_rate 2e-5 --weight_decay 0.1 --warmup_ratio 0.02 \
--max_steps 62 \
--log_interval 1 \
--gradient_checkpointing \
--bf16 \
--deepspeed /home/test/jd_train/ds_config_zero3.json \
--report_to none \
--eval_data_glob "/home/test/datasets/my_corpus/test.jsonl"