fsdp5_vanilla_speed.sh

CUDA_VISIBLE_DEVICES=4,5,6,7 torchrun --nnodes 1 --nproc_per_node 4 --master_port 29502 train_vanilla_speed.py \
    --batch-size 1 \
    --model_path "/AIRvePFS/ai4science/users/ai4science/users/zyk/seminat_backup/model/OLMo-2-0425-1B" \
    --decoder_layers 1 \
    --encoder_layers 1 \
    --mlp False \
    --position_embedding_type "absolute" \
    --base "pretrained" \
    --save_path "ckp/sft-v5-1e4-1sample-resume" \
    --save_name "sft-v5-1e4-1sample-resume" \
    --data_type SemiNATForMultiRoundMaskInputStream \
    --run_name "olmo2-1b-sft-v5-1sample" \
    --epochs 1 \
    --chunk_size_limit 5 \
    --save_steps 10000 \
    --lr 5e-4 \
    --weight_decay 0.1 \
    --eps 1e-8 \
    --dtype bf16 \
    --betas 0.9 0.95 \
    --warmup_ratio 0.1 \
    --data_path "/AIRvePFS/ai4science/users/ai4science/users/zyk/seminat/data/olmo2_base_prob_multi_0723_long_2w.jsonl" \
    --max_length 8192 \
    --data_processess_num 0 \
    --attn_implementation "flash_attention_2" \
    # --use_wandb \
    # --ptm_model_path "/AIRvePFS/ai4science/users/ai4science/users/zyk/seminat/ckp/pretrain/model.pt" \
    # --resume_path "/AIRvePFS/ai4science/users/ai4science/users/zyk/seminat/ckp/sft-v5-1e4-1sample-resume/sft-v5-1e4-1sample-resume-steps_300.pt" \