scripts/sft.sh

DATETIME=$(date '+%Y-%m-%d-%H')
RUN_NAME="timesearch-SFT-43-video-r1-data"
OUTPUT_DIR=/data/shuimu.chen/TimeSearch-R/experiment/$RUN_NAME/$DATETIME
mkdir -p $OUTPUT_DIR
export WANDB_PROJECT=TimeSearch-R-SFT
export WANDB_NAME=$RUN_NAME
export LOG_PATH=${OUTPUT_DIR}/log.txt
export DEBUG=true
export CUDA_VISIBLE_DEVICES=2,3,4,5,6,7

export NCCL_TIMEOUT=7200
export PYTHONPATH=".:$PYTHONPATH"
export SIGLIP_URL=grpc://127.0.0.1:51000
export WANDB_API_KEY="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"
# export LLM_AS_A_JUDGE_BASE=http://127.0.0.1:18901/v1

# Local training configuration
NUM_GPUS=6
MASTER_PORT=29500
echo "Local training mode: ${NUM_GPUS} GPUs on localhost:${MASTER_PORT}"


TRAIN_PATH=configs/dataset_sft.yaml

VIDEO_ROOT=/xuhongbo/shuimu.chen/LongVideoBench/videos_480p_noaudio


MODEL_BASE=/data/shuimu.chen/Qwen2.5-VL-7B-Instruct


torchrun --nproc_per_node=${NUM_GPUS} --nnodes=1 --node_rank=0 \
    --master_addr=localhost --master_port=${MASTER_PORT} \
    time_r1/sft.py \
    --deepspeed ./scripts/zero3.json \
    --output_dir $OUTPUT_DIR \
    --model_name_or_path $MODEL_BASE \
    --train_data_path $TRAIN_PATH \
    --video_folder $VIDEO_ROOT \
    --prompt_template v3 \
    --tool_name_list seek_video_frames \
    --total_video_tokens 10240 \
    --max_frames 60 \
    --min_per_frame_tokens 4 \
    --max_per_frame_tokens 192 \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 2 \
    --dataloader_num_workers 0 \
    --logging_steps 1 \
    --bf16 \
    --torch_dtype bfloat16 \
    --data_seed 42 \
    --gradient_checkpointing true \
    --attn_implementation flash_attention_2 \
    --learning_rate 1e-6 \
    --num_train_epochs 1 \
    --run_name $RUN_NAME \
    --report_to wandb \
    --save_steps 200 \
    --save_only_model true