请问目前是仅支持lora和qlora微调吗，全参数微调后续会开放吗？

[CUCldyyyyy]

谢谢各位作者优秀的工作！
请问各位author有没有做过实验评估全参数微调和peft后的模型性能各自怎么样呢？希望解惑～～

[CUCldyyyyy]

已经找到全参数微调的脚本了，运行时报错参数不匹配，请问是什么原因呢？
Traceback (most recent call last): File "/cfs-3xKdVIAptM/LDY/DB-GPT-Hub/dbgpt_hub/train/sft_train.py", line 165, in <module> train() File "/cfs-3xKdVIAptM/LDY/DB-GPT-Hub/dbgpt_hub/train/sft_train.py", line 142, in train run_sft( File "/cfs-3xKdVIAptM/LDY/DB-GPT-Hub/dbgpt_hub/train/sft_train.py", line 51, in run_sft dataset = preprocess_dataset(dataset, tokenizer, data_args, training_args, "sft") TypeError: preprocess_dataset() takes 4 positional arguments but 5 were given

这是我的脚本设置：

deepspeed --num_gpus 8 dbgpt_hub/train/sft_train.py
--dataset example_text2sql_train
--model_name_or_path /cfs-3xKdVIAptM/base_model/CodeLlama-13b-Instruct-hf
--do_train
--finetuning_type full
--max_source_length 2048
--max_target_length 512
--template llama2
--output_dir dbgpt_hub/output/adapter/CodeLlama-13b-sql-full
--overwrite_cache
--overwrite_output_dir
--per_device_train_batch_size 4
--gradient_accumulation_steps 16
--lr_scheduler_type cosine_with_restarts
--logging_steps 50
--learning_rate 2e-5
--num_train_epochs 4
--plot_loss
--bf16 false
--deepspeed dbgpt_hub/configs/ds_config.json 2>&1 | tee ${train_log}

[Kudou-Chitose]

已经找到全参数微调的脚本了，运行时报错参数不匹配，请问是什么原因呢？ Traceback (most recent call last): File "/cfs-3xKdVIAptM/LDY/DB-GPT-Hub/dbgpt_hub/train/sft_train.py", line 165, in <module> train() File "/cfs-3xKdVIAptM/LDY/DB-GPT-Hub/dbgpt_hub/train/sft_train.py", line 142, in train run_sft( File "/cfs-3xKdVIAptM/LDY/DB-GPT-Hub/dbgpt_hub/train/sft_train.py", line 51, in run_sft dataset = preprocess_dataset(dataset, tokenizer, data_args, training_args, "sft") TypeError: preprocess_dataset() takes 4 positional arguments but 5 were given

这是我的脚本设置：

deepspeed --num_gpus 8 dbgpt_hub/train/sft_train.py --dataset example_text2sql_train --model_name_or_path /cfs-3xKdVIAptM/base_model/CodeLlama-13b-Instruct-hf --do_train --finetuning_type full --max_source_length 2048 --max_target_length 512 --template llama2 --output_dir dbgpt_hub/output/adapter/CodeLlama-13b-sql-full --overwrite_cache --overwrite_output_dir --per_device_train_batch_size 4 --gradient_accumulation_steps 16 --lr_scheduler_type cosine_with_restarts --logging_steps 50 --learning_rate 2e-5 --num_train_epochs 4 --plot_loss --bf16 false --deepspeed dbgpt_hub/configs/ds_config.json 2>&1 | tee ${train_log}

今天碰到了同样的问题。检查后发现是前面跟着3.2 Quick Start执行了pip install dbgpt-hub命令，而pip版本里的preprocess_dataset()函数和现在的不一致。所以执行from dbgpt_hub.data_process.data_utils import xxx的时候实际上引用了旧的package。pip uninstall dbgpt-hub应该就可以了。

[CUCldyyyyy]

已经找到全参数微调的脚本了，运行时报错参数不匹配，请问是什么原因呢？ Traceback (most recent call last): File "/cfs-3xKdVIAptM/LDY/DB-GPT-Hub/dbgpt_hub/train/sft_train.py", line 165, in <module> train() File "/cfs-3xKdVIAptM/LDY/DB-GPT-Hub/dbgpt_hub/train/sft_train.py", line 142, in train run_sft( File "/cfs-3xKdVIAptM/LDY/DB-GPT-Hub/dbgpt_hub/train/sft_train.py", line 51, in run_sft dataset = preprocess_dataset(dataset, tokenizer, data_args, training_args, "sft") TypeError: preprocess_dataset() takes 4 positional arguments but 5 were given
这是我的脚本设置：
deepspeed --num_gpus 8 dbgpt_hub/train/sft_train.py --dataset example_text2sql_train --model_name_or_path /cfs-3xKdVIAptM/base_model/CodeLlama-13b-Instruct-hf --do_train --finetuning_type full --max_source_length 2048 --max_target_length 512 --template llama2 --output_dir dbgpt_hub/output/adapter/CodeLlama-13b-sql-full --overwrite_cache --overwrite_output_dir --per_device_train_batch_size 4 --gradient_accumulation_steps 16 --lr_scheduler_type cosine_with_restarts --logging_steps 50 --learning_rate 2e-5 --num_train_epochs 4 --plot_loss --bf16 false --deepspeed dbgpt_hub/configs/ds_config.json 2>&1 | tee ${train_log}

今天碰到了同样的问题。检查后发现是前面跟着3.2 Quick Start执行了pip install dbgpt-hub命令，而pip版本里的preprocess_dataset()函数和现在的不一致。所以执行from dbgpt_hub.data_process.data_utils import xxx的时候实际上引用了旧的package。pip uninstall dbgpt-hub应该就可以了。

谢谢🙏我当时把最后的sft参数注释掉了也work了:)