This commit is contained in:
@@ -0,0 +1,30 @@
|
||||
# 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 \
|
||||
--tuner_type full \
|
||||
--dataset swift/RLAIF-V-Dataset \
|
||||
--load_from_cache_file true \
|
||||
--split_dataset_ratio 0.01 \
|
||||
--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 \
|
||||
--rpo_alpha 0.1 \
|
||||
--sequence_parallel_size 4
|
||||
Reference in New Issue
Block a user