How do I set up GPU selection in a custom dataset fine-tuning script?

I have modified the script as follows:

(damo) [root@bgmm01 custom]# cat finetune.sh
#!/bin/bash

# Environment Variables
export CUDA_VISIBLE_DEVICES=4,5
WORLD_SIZE=${1:-1}
NPROC_PER_NODE=${2:-1} # Only use 1 process per node
MASTER_ADDR="127.0.0.1"
MASTER_PORT=16666
RANK=0

# Training Arguments
GLOBAL_BATCH_SIZE=128
LOCAL_BATCH_SIZE=4
GRADIENT_ACCUMULATION_STEPS=$[$GLOBAL_BATCH_SIZE/($WORLD_SIZE*$NPROC_PER_NODE*$LOCAL_BATCH_SIZE)]

# Log Arguments
export TRANSFORMERS_OFFLINE=1
export WANDB_PROJECT=videollama2_vllava
RUN_NAME=videollama2_vllava
DATA_DIR=datasets
OUTP_DIR=/data/VideoLLaMA2/work_dirs

torchrun --nnodes $WORLD_SIZE \
--nproc_per_node $NPROC_PER_NODE \
--master_addr=$MASTER_ADDR \
--master_port=$MASTER_PORT \
--node_rank $RANK \
/data/VideoLLaMA2/videollama2/train_flash_attn.py \
--deepspeed /data/VideoLLaMA2/scripts/zero3.json \
--version v1_mistral \
--vision_tower openai/clip-vit-large-patch14-336 \
--mm_projector_type stc_connector \
--model_name_or_path /data/VideoLLaMA2/Mistral-7B-Instruct-v0.2/Mistral-7B-Instruct-v0.2 \
--data_path /data/VideoLLaMA2/datasets/custom_sft/custom.json \
--data_folder /data/VideoLLaMA2/datasets/custom_sft/ \
--freeze_backbone True \
--pretrain_mm_mlp_adapter /data/VideoLLaMA2/model--base/mm_projector.bin \
--mm_vision_select_layer -2 \
--mm_use_im_start_end False \
--mm_use_im_patch_token False \
--image_aspect_ratio pad \
--num_frames 8 \
--bf16 True \
--tf32 True \
--fp16 False \
--output_dir ${OUTP_DIR}/${WANDB_PROJECT}/finetune_${RUN_NAME} \
--num_train_epochs 1 \
--per_device_train_batch_size $LOCAL_BATCH_SIZE \
--per_device_eval_batch_size 4 \
--gradient_accumulation_steps $GRADIENT_ACCUMULATION_STEPS \
--evaluation_strategy "no" \
--save_strategy "steps" \
--save_steps 500 \
--save_total_limit 99 \
--learning_rate 2e-5 \
--weight_decay 0. \
--warmup_ratio 0.03 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--model_max_length 2048 \
--gradient_checkpointing True \
--dataloader_num_workers 4 \
--report_to tensorboard \
--run_name $RUN_NAME

Generally, isn't it done by specifying GPUs with export CUDA_VISIBLE_DEVICES=4,5?

I have monitored, and these two GPUs are indeed almost completely idle, but when I run it, it prompts an error saying insufficient memory and it keeps using a GPU that is already running other models. How can this situation be resolved?

Read finetune.sh and the training entry point at videollama2/train_flash_attn.py. First compare the CUDA_VISIBLE_DEVICES values with the torchrun process settings and reproduce the launch; done means the training process uses only the requested GPUs and no longer reports memory pressure on another device.

Written by the indexing model from the issue text.