wehub-resource-sync
32 lines
999 B
Bash
32 lines
999 B
Bash
# Supported multimodal models reference:
|
|
# https://github.com/modelscope/ms-swift/blob/main/examples/train/packing/qwen2_5_vl.sh
|
|
# without padding_free: 4 * 60GiB, 26h
|
|
# padding_free: 4 * 44GiB, 13h
|
|
NPROC_PER_NODE=4 \
|
|
CUDA_VISIBLE_DEVICES=0,1,2,3 \
|
|
swift sft \
|
|
--model Qwen/Qwen2.5-7B \
|
|
--tuner_type full \
|
|
--dataset 'liucong/Chinese-DeepSeek-R1-Distill-data-110k-SFT' \
|
|
--load_from_cache_file true \
|
|
--split_dataset_ratio 0.01 \
|
|
--torch_dtype bfloat16 \
|
|
--per_device_train_batch_size 8 \
|
|
--per_device_eval_batch_size 8 \
|
|
--learning_rate 1e-5 \
|
|
--gradient_accumulation_steps 1 \
|
|
--eval_steps 200 \
|
|
--save_steps 200 \
|
|
--logging_steps 5 \
|
|
--max_length 8192 \
|
|
--warmup_ratio 0.05 \
|
|
--dataloader_num_workers 8 \
|
|
--dataset_num_proc 8 \
|
|
--save_total_limit 2 \
|
|
--save_only_model true \
|
|
--output_dir output/Qwen2.5-7B \
|
|
--deepspeed zero3 \
|
|
--use_liger_kernel true \
|
|
--attn_impl flash_attn \
|
|
--padding_free true
|