Upload folder using huggingface_hub

config.json

@@ -0,0 +1,52 @@
+{
+  "X": [
+    "Audio_asr",
+    "Audio_caption",
+    "Video",
+    "Image"
+  ],
+  "_name_or_path": "/mnt/bn/tns-algo-public-my2/wangpeng.an/model/Meta-Llama-3-8B-Instruct",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 128000,
+  "eos_token_id": 128009,
+  "freeze_mm_mlp_adapter": false,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "image_aspect_ratio": "square",
+  "image_grid_pinpoints": null,
+  "initializer_range": 0.02,
+  "intermediate_size": 14336,
+  "is_fusion": false,
+  "max_position_embeddings": 8192,
+  "mlp_bias": false,
+  "mm_audio_caption_tower": "/mnt/bn/tns-algo-public-my2/wangpeng.an/model/pretrained_model/LanguageBind_Audio",
+  "mm_audio_tower": "/mnt/bn/tns-algo-public-my2/wangpeng.an/model/LanguageBind_Audio_Asr",
+  "mm_hidden_size": 1024,
+  "mm_image_tower": "/mnt/bn/tns-algo-public-my2/wangpeng.an/model/pretrained_model/LanguageBind_Image",
+  "mm_projector_type": "mlp2x_gelu",
+  "mm_use_x_patch_token": false,
+  "mm_use_x_start_end": false,
+  "mm_video_tower": "/mnt/bn/tns-algo-public-my2/wangpeng.an/model/pretrained_model/LanguageBind_Video_merge",
+  "mm_vision_select_feature": "patch",
+  "mm_vision_select_layer": -2,
+  "model_type": "llava_llama",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 8,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 500000.0,
+  "tie_word_embeddings": false,
+  "tokenizer_model_max_length": 3072,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.43.1",
+  "tune_mm_mlp_adapter": true,
+  "use_cache": true,
+  "use_mm_proj": true,
+  "vocab_size": 128257
+}

mm_projector.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e57532ad31d36df8c8add40d8f2f7de3e27a009a19a266f71e1c44550341e2a6
+size 41961528

pretrain_all.sh

@@ -0,0 +1,91 @@
+#!/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
+export NCCL_SOCKET_FAMILY=AF_INET6
+export NCCL_NET_PLUGIN=libnccl-net-gcp-fastrak.so
+
+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-public-my2/wangpeng.an/environment/anaconda3/bin/activate multimodal
+
+cd /mnt/bn/tns-algo-public-my2/wangpeng.an/train/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 /mnt/bn/tns-algo-public-my2/wangpeng.an/model/Meta-Llama-3-8B-Instruct \
+    --version plain \
+    --data_path /mnt/bn/tns-algo-public-my2/wangpeng.an/data/annotations/video_image_caption_asr_stage1.json \
+    --audio_asr_folder /mnt/bn/tns-algo-public-my2/wangpeng.an/data \
+    --audio_caption_folder /mnt/bn/tns-algo-public-my2/wangpeng.an/data/caption_data/0818 \
+    --video_folder /mnt/bn/tns-algo-public-my2/wangpeng.an/data/video/Video-LLaVA \
+    --image_folder /mnt/bn/tns-algo-public-my2/wangpeng.an/data/video/Video-LLaVA \
+    --X "Audio_asr" "Audio_caption" "Video" "Image" \
+    --audio_tower /mnt/bn/tns-algo-public-my2/wangpeng.an/model/LanguageBind_Audio_Asr \
+    --audio_caption_tower /mnt/bn/tns-algo-public-my2/wangpeng.an/model/pretrained_model/LanguageBind_Audio \
+    --video_tower /mnt/bn/tns-algo-public-my2/wangpeng.an/model/pretrained_model/LanguageBind_Video_merge \
+    --image_tower /mnt/bn/tns-algo-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/Omni-Pretrain-8B-llama-0209 \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 4 \
+    --per_device_eval_batch_size 16 \
+    --gradient_accumulation_steps 2 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 2000 \
+    --save_total_limit 20 \
+    --learning_rate 1e-3 \
+    --weight_decay 0. \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --tf32 True \
+    --model_max_length 2048 \
+    --tokenizer_model_max_length 3072 \
+    --gradient_checkpointing True \
+    --dataloader_num_workers 8 \
+    --lazy_preprocess True \
+    --report_to tensorboard \
+    --cache_dir "./cache_dir"