Problem training on zero2.json

[Davidup1]

thank you for great job! but I met problem when training using zero2.json the loss and learning rate is 0.0 and i'm not sure how to fix it

[yaolug]

could you provide your full environment and command line

[Davidup1]

I set up the environment following the readme and my script is as follow

#!/bin/bash

export NCCL_P2P_DISABLE="1"
export NCCL_IB_DISABLE="1"

deepspeed --master_port=$((RANDOM + 10000)) --include localhost:0,1,2,3 /home/lyj/data/VILA/llava/train/train_mem.py \
    --deepspeed /home/lyj/data/VILA/scripts/zero2.json \
    --model_name_or_path /home/lyj/data/data/VILA1.5-3b/ \
    --version v1 \
    --data_mixture levir_cc \
    --vision_tower google/siglip-so400m-patch14-384 \
    --mm_vision_select_feature cls_patch \
    --mm_projector mlp_downsample \
    --tune_vision_tower False \
    --tune_mm_projector True \
    --tune_cc_projector True \
    --tune_single_projector True \
    --tune_language_model False \
    --mm_vision_select_layer -2 \
    --vision_resolution 576 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --image_aspect_ratio pad \
    --chg True \
    --chg_type Chg2Cap \
    --from_origin True \
    --cc_n_layers 3 \
    --cc_head 8 \
    --cc_dropout 0.1 \
    --bf16 True \
    --output_dir /home/lyj/data/data/checkpoints_1/ \
    --num_train_epochs 10 \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 8 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 0.5 \
    --save_total_limit 5 \
    --learning_rate 1e-4 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 4096 \
    --gradient_checkpointing True \
    --dataloader_num_workers 16 \
    --lazy_preprocess True \
    --vflan_no_system_prompt True \
    --report_to wandb

the script work well with zero3.json but went wrong with zero2.json

i added some modules and setting to the model structure for my specific tasks. The general modification is that I expanded the structure of mm_projector.

However, although I don’t know the specific reason, by modifying the import method of pre-training parameters, the model can now be trained normally under zero2.json. But I am still curious about the inducing mechanism of the problem.

[Lyken17]

This is ususally casued by specific transformer and deepspeed versions. Coud you run environment_setup.sh to reinstall the env?