InternLM-XComposer2-VL-7B使用lora微调，似乎保存了整个模型？

[handsomecaoyu]

我在用lora微调的时候，发现它保存的目录下有一个文件 mp_rank_00_model_states.pt，有32GB，还有一个bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt，有900M。有点困惑，lora应该只保存它训练的那部分参数才对。

我的finetune_lora.sh如下：

#!/bin/bash
export CUDA_DEVICE_MAX_CONNECTIONS=1
DIR=`pwd`


export MODEL="/root/onethingai-tmp/Shanghai_AI_Laboratory/internlm-xcomposer2-vl-7b"

export DATA="finetune/data.txt"

GPUS_PER_NODE=1
NNODES=1
NODE_RANK=0
MASTER_ADDR=localhost
MASTER_PORT=6001

DISTRIBUTED_ARGS="
    --nproc_per_node $GPUS_PER_NODE \
    --nnodes $NNODES \
    --node_rank $NODE_RANK \
    --master_addr $MASTER_ADDR \
    --master_port $MASTER_PORT
"

torchrun $DISTRIBUTED_ARGS finetune/finetune.py \
    --model_name_or_path $MODEL \
    --data_path $DATA \
    --img_size 490 \
    --given_num True \
    --bf16 True \
    --fix_vit True \
    --fix_sampler True \
    --use_lora True \
    --output_dir output/3_4_single_422_internlm \
    --num_train_epochs 20 \
    --batch_size 1 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 8 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 400 \
    --save_total_limit 10 \
    --learning_rate 5e-5 \
    --weight_decay 0.1 \
    --adam_beta2 0.95 \
    --warmup_ratio 0.01 \
    --lr_scheduler_type "cosine" \
    --logging_steps 10 \
    --report_to "none" \
    --max_length 4096 \
    --deepspeed finetune/ds_config_zero2.json \
    --gradient_checkpointing True

[yuhangzang]

For unwanted files, you can delete them manually.

[Starfulllll]

For unwanted files, you can delete them manually.

Excuse me, if I only save the parameters corresponding to the adapter (bin file), how should I load the trained model?

[yuhangzang]

For unwanted files, you can delete them manually.

Excuse me, if I only save the parameters corresponding to the adapter (bin file), how should I load the trained model?

See here