Upload folder using huggingface_hub

pretrain_all.sh

@@ -1,28 +1,75 @@
-CUDA_VISIBLE_DEVICES=0,1 deepspeed llava/train/train_mem.py \
+#!/bin/bash
+
+# Pretrain a multimodal model.
+export OMP_NUM_THREADS=8
+export NCCL_IB_DISABLE=0
+export NCCL_IB_GID_INDEX=3
+export NCCL_SOCKET_IFNAME=eth0
+export NCCL_DEBUG=INFO
+export NCCL_IB_SL=1
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+export NCCL_BLOCKING_WAIT=1
+export NCCL_ASYNC_ERROR_HANDLING=1
+export NCCL_TIMEOUT=500
+export TORCH_DISTRIBUTED_DEBUG=DETAIL
+
+DATETIME=`date +'%y-%m-%d-%H-%M-%S'`
+
+# Setting for multi nodes training.
+ports=(`echo $METIS_WORKER_0_PORT | tr ',' ' '`)
+port=${ports[0]}
+
+echo "total workers: ${ARNOLD_WORKER_NUM}"
+echo "cur worker id: ${ARNOLD_ID}"
+echo "gpus per worker: ${ARNOLD_WORKER_GPU}"
+echo "master ip: ${METIS_WORKER_0_HOST}"
+echo "master port: ${port}"
+
+source /mnt/bn/tns-algo-video-public-my2/wangpeng.an/environment/anaconda3/bin/activate multimodal
+
+cd /mnt/bn/tns-algo-video-public-my2/wangpeng.an/model/OmniFusion-main
+
+# Install necessary packages
+pip3 install requests
+pip3 install attrs
+pip3 install aiofiles
+pip3 install pynvml
+
+
+# Print Python executable path, torchrun, deepspeed and PYTHONPATH
+echo "Python executable: $(which python)"
+echo "torchrun executable: $(which torchrun)"
+echo "deepspeed executable: $(which deepspeed)"
+echo "PYTHONPATH before torchrun: $PYTHONPATH"
+
+sudo chmod 777 /var/lib/fastrak -R
+
+ACCELERATE_CPU_AFFINITY=1 torchrun --nproc_per_node=$ARNOLD_WORKER_GPU --nnodes=$ARNOLD_WORKER_NUM --node_rank=$ARNOLD_ID --master_addr=$METIS_WORKER_0_HOST --master_port=$port \
+    llava/train/train_mem.py \
     --deepspeed ./scripts/zero2.json \
-    --model_name_or_path ./Meta-Llama-3.1-8B-Instruct \
+    --model_name_or_path /mnt/bn/tns-algo-video-public-my2/wangpeng.an/model/Meta-Llama-3.1-8B-Instruct \
     --version plain \
-    --data_path train_json/video_image_asr_caption_pre.json \
-    --audio_asr_folder /home/yu/Downloads \
-    --audio_caption_folder /media/yu/33da5c9f-d06a-4b18-acc4-b4e1c20292471/0818 \
-    --video_folder video_data \
-    --image_folder /media/yu/33da5c9f-d06a-4b18-acc4-b4e1c20292471 \
+    --data_path /mnt/bn/tns-algo-video-public-my2/wangpeng.an/data/multidataset/video_image_asr_caption_pre_1102.json \
+    --audio_asr_folder /mnt/bn/tns-algo-video-public-my2/wangpeng.an/data \
+    --audio_caption_folder /mnt/bn/tns-algo-video-public-my2/wangpeng.an/data/multidataset/caption_data/0818 \
+    --video_folder /mnt/bn/tns-algo-video-public-my2/wangpeng.an/data/Video-LLaVA \
+    --image_folder /mnt/bn/tns-algo-video-public-my2/wangpeng.an/data/Video-LLaVA \
     --X "Audio_asr" "Audio_caption" "Video" "Image" \
-    --audio_tower ./LanguageBind_Audio_Asr \
-    --audio_caption_tower LanguageBind/LanguageBind_Audio \
-    --video_tower LanguageBind/LanguageBind_Video_merge \
-    --image_tower LanguageBind/LanguageBind_Image \
+    --audio_tower /mnt/bn/tns-algo-video-public-my2/wangpeng.an/model/LanguageBind_Audio_Asr \
+    --audio_caption_tower /mnt/bn/tns-algo-video-public-my2/wangpeng.an/model/pretrained_model/LanguageBind_Audio \
+    --video_tower /mnt/bn/tns-algo-video-public-my2/wangpeng.an/model/pretrained_model/LanguageBind_Video_merge \
+    --image_tower /mnt/bn/tns-algo-video-public-my2/wangpeng.an/model/pretrained_model/LanguageBind_Image \
     --mm_projector_type mlp2x_gelu \
     --tune_mm_mlp_adapter True \
     --mm_vision_select_layer -2 \
     --mm_use_x_start_end False \
     --mm_use_x_patch_token False \
     --bf16 True \
-    --output_dir ./checkpoints/Video-LLaVA-Pretrain-7B \
+    --output_dir ./checkpoints/Video-LLaVA-Pretrain-7B-1102 \
     --num_train_epochs 1 \
-    --per_device_train_batch_size 1 \
+    --per_device_train_batch_size 16 \
     --per_device_eval_batch_size 16 \
-    --gradient_accumulation_steps 32 \
+    --gradient_accumulation_steps 1 \
     --evaluation_strategy "no" \
     --save_strategy "steps" \
     --save_steps 2000 \