sft_8frames.sh

#!/bin/bash
# Set default values for distributed training parameters
# 禁用WandB
export WANDB_DISABLED=true
# ===== GPU配置参数 =====
# export CUDA_VISIBLE_DEVICES=0,1,2,3  # 使用4张GPU
export CUDA_VISIBLE_DEVICES=0  # GPU编号
export GPUS_PER_NODE=1               # 每个节点GPU数量
export n_node=1                      # 单节点
    export CURRENT_RANK=0                # 当前节点rank
    export MASTER_ADDR="localhost"       # 主节点地址
    export MASTER_PORT=48057             # 主节点端口

llama3_model_path="/home/user/NaVILA/checkpoints/navila-llama3-8b-8f"
GEOMETRY_ENCODER_TYPE="pi3"
GEOMETRY_ENCODER_PATH="/home/user/NaVILA/checkpoints/Pi3"
# GEOMETRY_ENCODER_TYPE="vggt"
# GEOMETRY_ENCODER_PATH="/home/user/NaVILA/checkpoints/VGGT-1B"

OUTPUT="/home/user/NaVILA/sft_8frames"

# 检查是否有检查点可以resume
if [ -d "$OUTPUT" ] && [ "$(ls -A $OUTPUT | grep -E 'tmp-checkpoint-|checkpoint-')" ]; then
    echo "Found existing checkpoints, will resume from the latest one"
    # 找到最新的检查点
    LATEST_CHECKPOINT=$(ls -t $OUTPUT | grep -E 'tmp-checkpoint-|checkpoint-' | head -1)
    RESUME_PATH="$OUTPUT/$LATEST_CHECKPOINT"
    echo "Resuming from: $RESUME_PATH"
    # 不删除输出目录，让训练自动resume
else
    echo "No existing checkpoints found, starting fresh training"
    # rm -rf /home/user/NaVILA/output*
    RESUME_PATH="$llama3_model_path"
fi

torchrun --nnodes=$n_node --nproc_per_node=$GPUS_PER_NODE --master_port=$MASTER_PORT \
    --master_addr $MASTER_ADDR --node_rank=$CURRENT_RANK \
    llava/train/train_mem.py \
    --longvila_sampler True \
    --deepspeed ./scripts/zero3.json \
    --model_name_or_path $RESUME_PATH \
    --version llama_3 \
    --seed 3407 \
    --data_mixture r2r+rxr \
    --vision_tower google/siglip-so400m-patch14-384 \
    --mm_vision_select_feature cls_patch \
    --mm_projector mlp_downsample \
    --num_video_frames 16 \
    --tune_vision_tower False \
    --tune_mm_projector True \
    --tune_language_model True \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --image_aspect_ratio resize \
    --bf16 True \
    --output_dir $OUTPUT \
    --num_train_epochs 5 \
    --per_device_train_batch_size 10 \
    --gradient_accumulation_steps 2 \
    --do_eval False \
    --save_strategy "steps" \
    --save_steps 2000 \
    --fps 0.0 \
    --save_total_limit 10 \
    --learning_rate 1e-4 \
    --mm_projector_lr 1e-5 \
    --geometry_merger_lr 1e-4 \
    --feature_fusion_lr 1e-4 \
    --llm_lr 1e-6 \
    --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 \
    --report_to tensorboard \
    --logging_dir $OUTPUT/runs \
    --use_geometry_encoder True \
    --geometry_encoder_type $GEOMETRY_ENCODER_TYPE \
    --geometry_encoder_path $GEOMETRY_ENCODER_PATH \
    --feature_fusion_method "residual" \
    --voxel_size 0.2 \