How to go on fine-tuning with a lora checkpoint?

[ORGRUI]

I tried using resume_from_checkpoint, but I got an error.

tag: v0.0.8

script:
deepspeed ${deepspeed_args}
examples/finetune.py
--model_name_or_path ${model_name_or_path}
--trust_remote_code ${trust_remote_code}
--dataset_path ${dataset_path}
--conversation_template ${conversation_template}
--output_dir ${output_dir} --overwrite_output_dir
--resume_from_checkpoint ${resume_from_checkpoint_path}
--num_train_epochs 8
--learning_rate 2e-4
--block_size 1024
--per_device_train_batch_size 2
--use_qlora 1
--save_aggregated_lora 0
--deepspeed configs/ds_config_zero2.json
--fp16
--run_name ${exp_id}
--validation_split_percentage 0
--logging_steps 20
--do_train
--ddp_timeout 72000
--save_steps 200
--dataloader_num_workers 1
| tee ${log_dir}/train.log
2> ${log_dir}/train.err

[rank0]: Traceback (most recent call last):
[rank0]: File "/data/rio/LMFlow-main/examples/finetune.py", line 61, in
[rank0]: main()
[rank0]: File "/data/rio/LMFlow-main/examples/finetune.py", line 57, in main
[rank0]: tuned_model = finetuner.tune(model=model, dataset=dataset)
[rank0]: File "/data/alex/LMFlow/src/lmflow/pipeline/finetuner.py", line 591, in tune
[rank0]: train_result = trainer.train(resume_from_checkpoint=checkpoint)
[rank0]: File "/home/robotuser/miniconda3/envs/lmflow_rio/lib/python3.9/site-packages/transformers/trainer.py", line 1938, in train
[rank0]: return inner_training_loop(
[rank0]: File "/home/robotuser/miniconda3/envs/lmflow_rio/lib/python3.9/site-packages/transformers/trainer.py", line 2119, in _inner_training_loop
[rank0]: deepspeed_load_checkpoint(
[rank0]: File "/home/robotuser/miniconda3/envs/lmflow_rio/lib/python3.9/site-packages/transformers/integrations/deepspeed.py", line 442, in deepspeed_load_checkpoint
[rank0]: raise ValueError(f"Can't find a valid checkpoint at {checkpoint_path}")
[rank0]: ValueError: Can't find a valid checkpoint at /data/rio/LMFlow-main/output_models/opening_sft_llama3_70b_1223/result/adapter_model
[2024-12-31 18:04:48,756] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 4143052
[2024-12-31 18:04:48,757] [ERROR] [launch.py:322:sigkill_handler] ['/home/robotuser/miniconda3/envs/lmflow_rio/bin/python', '-u', 'examples/finetune.py', '--local_rank=0', '--model_name_or_path', '/data/hf_cache/hub/models--meta-llama--Meta-Llama-3-70B-Instruct/snapshots/5fcb2901844dde3111159f24205b71c25900ffbd', '--trust_remote_code', '0', '--dataset_path', '/data/rio/LMFlow-main/data/opening_10000_views', '--conversation_template', 'llama3', '--output_dir', '/data/rio/LMFlow-main/output_models/opening_10000_views', '--overwrite_output_dir', '--resume_from_checkpoint', '/data/rio/LMFlow-main/output_models/opening_sft_llama3_70b_1223/result/adapter_model', '--num_train_epochs', '8', '--learning_rate', '2e-4', '--block_size', '1024', '--per_device_train_batch_size', '2', '--use_qlora', '1', '--save_aggregated_lora', '0', '--deepspeed', 'configs/ds_config_zero2.json', '--fp16', '--run_name', 'finetune_with_qlora_opening_10000_views', '--validation_split_percentage', '0', '--logging_steps', '20', '--do_train', '--ddp_timeout', '72000', '--save_steps', '200', '--dataloader_num_workers', '1'] exits with return code = 1

[wheresmyhair]

Hi, thank you for your interest in LMFlow! Please specify the trained LoRA adapter using lora_model_path instead of resume_from_checkpoint. We will clarify the documentation further in the future. Apologies for any confusion!