[Question]: Training S2 System StreamVLN

[kason-huang]

Question

Hello, I tried to reproduce the result of StreamVLN， and I use an H20 machine with 96GB of VRAM, but I sitll encountered a cuda out of memory error, and my configuration is

torchrun --nnodes=$NNODES --nproc_per_node=$NPROC_PER_NODE \
    --rdzv_id=12345 --rdzv_backend=c10d --rdzv_endpoint=$MASTER_ADDR:$MASTER_PORT streamvln/streamvln_train.py \
    --deepspeed scripts/zero2.json \
    --model_name_or_path $PREV_STAGE_CHECKPOINT \
    --version $PROMPT_VERSION \
    --video_folder ${VIDEO_FOLDER} \
    --group_by_task False \
    \
    --num_history 8 \
    --num_future_steps 4 \
    --num_frames 32 \
    --data_augmentation True \
    \
    --mm_tunable_parts="mm_vision_tower,mm_mlp_adapter,mm_language_model" \
    --vision_tower ${VISION_MODEL_VERSION} \
    --mm_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --image_aspect_ratio anyres_max_9 \
    --image_grid_pinpoints  "(1x1),...,(6x6)" \
    --bf16 True \
    --run_name $MID_RUN_NAME \
    --output_dir checkpoints/$MID_RUN_NAME \
    --num_train_epochs 1 \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 2 \
    --evaluation_strategy "no" \
    --save_strategy "epoch" \
    --save_total_limit 1 \
    --learning_rate 2e-5 \
    --mm_vision_tower_lr 5e-6 \
    --weight_decay 0. \
    --warmup_ratio 0.075 \
    --lr_scheduler_type "cosine_with_min_lr" \
    --lr_scheduler_kwargs '{"min_lr": 1.85e-05}' \
    --logging_steps 10 \
    --tf32 True \
    --model_max_length 32768 \
    --gradient_checkpointing True \
    --dataloader_num_workers 8 \
    --lazy_preprocess True \
    --torch_compile True \
    --torch_compile_backend "inductor" \
    --dataloader_drop_last True \
    --report_to tensorboard

However, modifying mm_tunable_parts=mm_vision_tower,mm_mlp_adapter resolved the issue. I'd like to ask is the model performance described in the paper was achived with mm_tunable_parts="mm_vision_tower,mm_mlp_adapter,mm_language_model" or no?
And by the way, is the dataset EnvDrop dataset used in the following experiment？

[kellyiss]

1. Yes, mm_tunable_parts="mm_vision_tower,mm_mlp_adapter,mm_language_model".
2. You can choose to use EnvDrop dataset or just train on r2r dataset and rxr dataset for more epochs.