train loss 0 with vl

[betterftr]

When I'm training internlm-xcomposer2-7b:

When I'm training internlm-xcomposer2-vl-7b:

[bailanwa]

I also encountered the same problem

[yuhangzang]

This issue has been fixed by the commit a57c093. You can pull the latest version of fine-tune code to check if the problem still exists.

[jpWang]

Hi, When I use 2 machines with 8xA100 for each, the loss still becomes 0:

{'loss': 1.7172, 'learning_rate': 7.874015748031497e-08, 'epoch': 0.0}
{'loss': 1.7636, 'learning_rate': 1.5748031496062994e-07, 'epoch': 0.0}
{'loss': 1.7636, 'learning_rate': 1.5748031496062994e-07, 'epoch': 0.0}
{'loss': 1.7367, 'learning_rate': 2.362204724409449e-07, 'epoch': 0.0}
{'loss': 1.7367, 'learning_rate': 2.362204724409449e-07, 'epoch': 0.0}
{'loss': 1.7638, 'learning_rate': 3.149606299212599e-07, 'epoch': 0.0}
{'loss': 1.7638, 'learning_rate': 3.149606299212599e-07, 'epoch': 0.0}
{'loss': 0.0, 'learning_rate': 3.937007874015748e-07, 'epoch': 0.0}
{'loss': 0.0, 'learning_rate': 3.937007874015748e-07, 'epoch': 0.0}
{'loss': 0.0, 'learning_rate': 4.724409448818898e-07, 'epoch': 0.0}
{'loss': 0.0, 'learning_rate': 4.724409448818898e-07, 'epoch': 0.0}
{'loss': 0.0, 'learning_rate': 5.511811023622048e-07, 'epoch': 0.0}
{'loss': 0.0, 'learning_rate': 5.511811023622048e-07, 'epoch': 0.0}

I use this to train:

#!/bin/bash
export NCCL_DEBUG=INFO
export NCCL_P2P_DISABLE=1
export NCCL_IB_DISABLE=1
export NCCL_SOCKET_IFNAME=eth0
DIR=`pwd`
GPUS_PER_NODE=$(python -c 'import torch; print(torch.cuda.device_count())')

DISTRIBUTED_ARGS="
    --nproc_per_node $GPUS_PER_NODE \
    --nnodes $WORLD_SIZE \
    --node_rank $RANK \
    --master_addr $MASTER_ADDR \
    --master_port $MASTER_PORT
"
torchrun $DISTRIBUTED_ARGS finetune.py \
    --model_name_or_path ./internlm-xcomposer2-vl-7b \
    --data_path $DATA \
    --img_size 490 \
    --bf16 True \
    --fix_vit False \
    --fix_sampler False \
    --use_lora False \
    --output_dir output/xxx \
    --num_train_epochs 1 \
    --batch_size 1 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 4 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 500 \
    --save_total_limit 15 \
    --learning_rate 1e-5 \
    --weight_decay 0.1 \
    --adam_beta2 0.95 \
    --warmup_ratio 0.01 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --report_to "none" \
    --max_length 4096 \
    --deepspeed ds_config_zero2.json \
    --gradient_checkpointing True

[jpWang]

@yuhangzang @myownskyW7 Looking forward to your reply, thank you！

[Sanster]

This is a bug related to deepspeed zero2, you can refer to this issue: https://github.com/haotian-liu/LLaVA/issues/1231