README_mine.md

Conda env (agilab2)

• L40S 40G * 8대

conda activate emu3_dpo
pip uninstall -y deepspeed && pip install deepspeed==0.15.4

TODO

• Clone repository.
• Prepare data with HumanEdit dataset.
• Train official code (SFT) with HumanEdit dataset.
• Eval SFT Model for MagicBrush (Edit) train/test dataset (L1, CLIP-T)
• Eval Original Model for MagicBrush (Edit) train/test dataset (L1, CLIP-T)
• Implement DPOTrainer with Emu3-Gen.
• Train DPO code with HumanEdit dataset.

[issue 1] OOM in Emu3 PEFT Training

(오피셜 코드 대비 수정사항 기록)

1. t2i_sft_offload.sh

• export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
• --image_area 262144
• --max_position_embeddings 5120
• --per_device_train_batch_size 1
• --gradient_accumulation_steps 8

2. zero3_offload.json

• sub_group_size
• reduce_bucket_size
• stage3_prefetch_bucket_size
• stage3_max_live_parameters
• stage3_max_reuse_distance
• pin_memory

3. train.py

1. add LoRA adapter