wehub-resource-sync
51 lines
1.4 KiB
Bash
51 lines
1.4 KiB
Bash
|
|
export teacher_model='Qwen/Qwen3-8B'
|
|
|
|
NPROC_PER_NODE=4 \
|
|
CUDA_VISIBLE_DEVICES=0,1,2,3 \
|
|
swift infer \
|
|
--model $teacher_model \
|
|
--infer_backend vllm \
|
|
--val_dataset 'AI-ModelScope/alpaca-gpt4-data-en#5000' 'AI-ModelScope/alpaca-gpt4-data-zh#5000' \
|
|
--vllm_gpu_memory_utilization 0.9 \
|
|
--vllm_max_model_len 8192 \
|
|
--max_new_tokens 2048 \
|
|
--write_batch_size 10000 \
|
|
--result_path new_dataset.jsonl
|
|
|
|
|
|
# 4 * 67GiB, 2.50s/it
|
|
# You need to additionally add sft_loss, because tokens like '<think>' have not been trained.
|
|
NPROC_PER_NODE=4 \
|
|
PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
|
|
CUDA_VISIBLE_DEVICES=0,1,2,3 \
|
|
swift rlhf \
|
|
--rlhf_type gkd \
|
|
--model Qwen/Qwen3-8B-Base \
|
|
--teacher_model $teacher_model \
|
|
--tuner_type full \
|
|
--dataset 'new_dataset.jsonl' \
|
|
--load_from_cache_file true \
|
|
--split_dataset_ratio 0.01 \
|
|
--torch_dtype bfloat16 \
|
|
--num_train_epochs 1 \
|
|
--learning_rate 1e-5 \
|
|
--per_device_train_batch_size 1 \
|
|
--per_device_eval_batch_size 1 \
|
|
--gradient_accumulation_steps 1 \
|
|
--eval_steps 100 \
|
|
--save_steps 100 \
|
|
--save_total_limit 2 \
|
|
--logging_steps 5 \
|
|
--max_length 4096 \
|
|
--output_dir output \
|
|
--warmup_ratio 0.05 \
|
|
--save_only_model true \
|
|
--dataloader_num_workers 4 \
|
|
--dataset_num_proc 4 \
|
|
--deepspeed zero3 \
|
|
--packing true \
|
|
--attn_impl flash_attn \
|
|
--sft_alpha 0.1 \
|
|
--lmbda 0
|