# pip install flash-attn --no-build-isolation


nproc_per_node=1

INFONCE_USE_BATCH=False \
CUDA_VISIBLE_DEVICES=0 \
NPROC_PER_NODE=$nproc_per_node \
swift sft \
    --model $(pwd)/../../data/models/Qwen3-Embedding-0.6B/model \
    --task_type embedding \
    --model_type qwen3_emb \
    --train_type lora \
    --lora_rank 8 \
    --lora_alpha 16 \
    --target_modules all-linear \
    --dataset my_local_dataset \
    --custom_register_path $(pwd)/../../data/dataset/my_dataset_register.py \
    --split_dataset_ratio 0.005 \
    --eval_strategy steps \
    --output_dir output \
    --eval_steps 1000 \
    --num_train_epochs 1 \
    --save_steps 1000 \
    --save_total_limit 10 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 16 \
    --gradient_accumulation_steps 4 \
    --learning_rate 2.4e-5 \
    --loss_type infonce \
    --label_names labels \
    --dataloader_drop_last true \
    --deepspeed zero3