File size: 796 Bytes
cb2428f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 |
# Env: 4 * A100
# GPU Memory: 4 * 25GiB, Training Speed 14s/it
NPROC_PER_NODE=4 \
CUDA_VISIBLE_DEVICES=0,1,2,3 \
swift rlhf \
--rlhf_type dpo \
--model Qwen/Qwen2.5-VL-3B-Instruct \
--train_type full \
--dataset swift/RLAIF-V-Dataset \
--torch_dtype bfloat16 \
--per_device_train_batch_size 1 \
--per_device_eval_batch_size 1 \
--learning_rate 1e-5 \
--gradient_accumulation_steps 8 \
--eval_steps 200 \
--save_steps 200 \
--logging_steps 5 \
--warmup_ratio 0.05 \
--dataloader_num_workers 8 \
--dataset_num_proc 8 \
--save_total_limit 2 \
--save_only_model true \
--output_dir output/Qwen2.5-VL-3B-Instruct \
--deepspeed zero3 \
--attn_impl flash_attn \
--use_liger_kernel true \
--sequence_parallel_size 4
|