PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True accelerate launch --config_file "configs/fsdp_config.yaml" \
--main_process_ip "172.51.185.4" \
--main_process_port "29080" \
--machine_rank $BACKENDAI_CLUSTER_LOCAL_RANK \
--num_processes 8 \
--num_machines 2 \
--monitor_interval 10 \
--num_cpu_threads_per_process 8 \
train.py \
--seed 100 \
--model_name_or_path "meta-llama/Meta-Llama-3.1-8B-Instruct" \
--output_dir "results/mncai/Meta-Llama-3.1-8B-Instruct-v0.2" \
--dataset_name "mncai/ai2_arc,mncai/hellaswag,mncai/mbpp,mncai/openbookqa,mncai/winogrande,mncai/trivia_qa,mncai/KMMLU" \
--chat_template_format "llama3.1" \
--add_special_tokens False \
--append_concat_token False \
--splits "train,test" \
--max_seq_len 8192 \
--num_train_epochs 1 \
--logging_steps 5 \
--log_level "info" \
--logging_strategy "steps" \
--eval_strategy "steps" \
--eval_steps 200 \
--save_total_limit 2 \
--bf16 True \
--packing False \
--learning_rate 1e-4 \
--lr_scheduler_type "cosine" \
--weight_decay 1e-4 \
--warmup_ratio 0.0 \
--max_grad_norm 1.0 \
--per_device_train_batch_size 8 \
--per_device_eval_batch_size 8 \
--gradient_accumulation_steps 1 \
--gradient_checkpointing True \
--use_reentrant False \
--dataset_text_field "content" \
--use_flash_attn True \
--optim paged_adamw_32bit \
--report_to "wandb" \
# --hub_model_id "mncai/Meta-Llama-3.1-8B-Instruct-v0.2" \
# --push_to_hub True \
# --hub_private_repo True \
# --hub_strategy "end" \
# --hub_token $HF_WRITE_TOKEN \

# --save_strategy "steps" \
# --save_steps 200 \
# --use_peft_lora True \
# --lora_r 8 \
# --lora_alpha 16 \
# --lora_dropout 0.1 \
# --lora_target_modules "all-linear" \
# --use_4bit_quantization False