Skip to content

Latest commit

 

History

History
196 lines (163 loc) · 6 KB

Bunny-v1.1-Llama-3-8B-V.md

File metadata and controls

196 lines (163 loc) · 6 KB
Raw

Training Tutorial of Bunny-v1.1-Llama-3-8B-V

Bunny-v1.1-Llama-3-8B-V

Pretrain

#!/bin/bash

MODEL_TYPE=llama3-8b
OUTPUT_DIR=bunny-$MODEL_TYPE-pretrain

mkdir -p ./checkpoints-pretrain/$OUTPUT_DIR

deepspeed bunny/train/train.py \
    --deepspeed ./script/deepspeed/zero3.json \
    --model_name_or_path /path/to/meta-llama/Meta-Llama-3-8B-Instruct \
    --model_type $MODEL_TYPE \
    --version plain \
    --data_path ./data/pretrain/bunny_pretrain_laion_2m.json \
    --image_folder ./data/pretrain/images \
    --vision_tower /path/to/siglip-so400m-patch14-384 \
    --use_s2 True \
    --mm_projector_type mlp2x_gelu \
    --tune_mm_mlp_adapter True \
    --image_aspect_ratio square \
    --bf16 True \
    --output_dir ./checkpoints-pretrain/$OUTPUT_DIR \
    --num_train_epochs 1 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 24000 \
    --save_total_limit 1 \
    --learning_rate 1e-3 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --report_to none | tee 2>&1 ./checkpoints-pretrain/$OUTPUT_DIR/log.txt

Visual Instruction Tuning

Recipe-1

#!/bin/bash

MODEL_TYPE=llama3-8b

PRETRAIN_DIR=bunny-$MODEL_TYPE-pretrain
OUTPUT_DIR=bunny-lora-$MODEL_TYPE-recipe-1

mkdir -p ./checkpoints-$MODEL_TYPE/$OUTPUT_DIR

deepspeed bunny/train/train.py \
    --lora_enable True --lora_r 256 --lora_alpha 256 --mm_projector_lr 1e-5 \
    --deepspeed ./script/deepspeed/zero3.json \
    --model_name_or_path /path/to/meta-llama/Meta-Llama-3-8B-Instruct \
    --model_type $MODEL_TYPE \
    --version llama \
    --data_path ./data/finetune/bunny_allava_1.3m.json \
    --image_folder ./data/finetune/images \
    --vision_tower /path/to/siglip-so400m-patch14-384 \
    --use_s2 True \
    --pretrain_mm_mlp_adapter ./checkpoints-pretrain/$PRETRAIN_DIR/mm_projector.bin \
    --mm_projector_type mlp2x_gelu \
    --image_aspect_ratio pad \
    --group_by_modality_length False \
    --bf16 True \
    --output_dir ./checkpoints-$MODEL_TYPE/$OUTPUT_DIR \
    --num_train_epochs 1 \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 500 \
    --save_total_limit 1 \
    --learning_rate 1e-4 \
    --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 4 \
    --lazy_preprocess True \
    --report_to none | tee 2>&1 ./checkpoints-$MODEL_TYPE/$OUTPUT_DIR/log.txt

Recipe-2

#!/bin/bash

MODEL_TYPE=llama3-8b

PRETRAIN_DIR=bunny-$MODEL_TYPE-pretrain
OUTPUT_DIR=bunny-lora-$MODEL_TYPE-recipe-2

mkdir -p ./checkpoints-$MODEL_TYPE/$OUTPUT_DIR

deepspeed bunny/train/train.py \
    --lora_enable True --lora_r 256 --lora_alpha 256 --mm_projector_lr 1e-5 \
    --deepspeed ./script/deepspeed/zero3.json \
    --model_name_or_path /path/to/meta-llama/Meta-Llama-3-8B-Instruct \
    --model_type $MODEL_TYPE \
    --version llama \
    --data_path ./data/finetune/bunny_llava_allava_2m.json \
    --image_folder ./data/finetune/images \
    --vision_tower /path/to/siglip-so400m-patch14-384 \
    --use_s2 True \
    --unfreeze_vision_tower True \
    --pretrain_mm_mlp_adapter ./checkpoints-pretrain/$PRETRAIN_DIR/mm_projector.bin \
    --mm_projector_type mlp2x_gelu \
    --image_aspect_ratio pad \
    --group_by_modality_length False \
    --bf16 True \
    --output_dir ./checkpoints-$MODEL_TYPE/$OUTPUT_DIR \
    --num_train_epochs 1 \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 500 \
    --save_total_limit 1 \
    --learning_rate 1e-4 \
    --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 4 \
    --lazy_preprocess True \
    --report_to none | tee 2>&1 ./checkpoints-$MODEL_TYPE/$OUTPUT_DIR/log.txt

Weight Merging

  • Firstly, merge the LoRA weights and base LLM

    python script/merge_lora_weights.py \
  --model-path ./checkpoints-llama3-8b/bunny-lora-llama3-8b-recipe-1 \
  --model-base /path/to/meta-llama/Meta-Llama-3-8B-Instruct \
  --model-type llama3-8b \
  --save-model-path ./checkpoints-llama3-8b/bunny-llama3-8b-recipe-1
    python script/merge_lora_weights.py \
  --model-path ./checkpoints-llama3-8b/bunny-lora-llama3-8b-recipe-2 \
  --model-base /path/to/meta-llama/Meta-Llama-3-8B-Instruct \
  --model-type llama3-8b \
  --save-model-path ./checkpoints-llama3-8b/bunny-llama3-8b-recipe-2

  • Then, inherit configurations from recipe-2

    cp -r ./checkpoints-llama3-8b/bunny-llama3-8b-recipe-2 ./checkpoints-llama3-8b/bunny-llama3-8b-avg

  • Lastly, linearly avearge two models

    from safetensors.torch import load_file, save_file

total = 4
for i in range(1, total + 1):
    model_1 = load_file(f'./checkpoints-llama3-8b/bunny-lora-llama3-8b-recipe-1/model-{i:05d}-of-{total:05d}.safetensors')
    model_2 = load_file(f'./checkpoints-llama3-8b/bunny-lora-llama3-8b-recipe-2/model-{i:05d}-of-{total:05d}.safetensors')
    
    assert model_1.keys() == model_2.keys()

    avg = {}
    for k in model_1.keys():
        avg[k] = model_1[k] * 0.42 + model_2[k] * 0.58 # the weight factor is selected empirically

    save_file(avg, f'./checkpoints-llama3-8b/bunny-llama3-8b-avg/model-{i:05d}-of-{total:05d}.safetensors', {'format': 'pt'})