official code lead to OOM !

[Xuekai-Zhu]

With transformers=3.31.0 and accelerate=0.22.0.dev0 (latest main) On 2 x A6000 48G, fine-tuning LLaMA 7B , the warning is:

FutureWarning: using `--fsdp_transformer_layer_cls_to_wrap` is deprecated. Use fsdp_config instead
FSDP Warning: When using FSDP, it is efficient and recommended to call prepare for the model before creating the optimizer.
FSDP Warning: When using FSDP, several parameter groups will be conflated into a single one due to nested module wrapping and parameter flattening.

And i use the default command:

torchrun --nproc_per_node=2 --master_port=20001 fastchat/train/train_mem.py \
    --model_name_or_path /root/zhuxuekai/data_pruning/pre-trained-model/huggyllama/llama-7b  \
    --data_path data/dummy_conversation.json \
    --bf16 True \
    --output_dir output_vicuna \
    --num_train_epochs 3 \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 2 \
    --gradient_accumulation_steps 16 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 1200 \
    --save_total_limit 10 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --fsdp "full_shard auto_wrap" \
    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
    --tf32 True \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --lazy_preprocess True \
    --disable_tqdm False

I install fastchat from source.

And i also try your solution, it is also not worked.

I found you have the problem, please help me to fix this !

[Xuekai-Zhu]

I use this hacking method to revise the code in trainer.py ; the warning is disappeared but still hit OOM;

Still warning:

FutureWarning: using `--fsdp_transformer_layer_cls_to_wrap` is deprecated. Use fsdp_config instead

the following warning disappear:

FSDP Warning: When using FSDP, it is efficient and recommended to call prepare for the model before creating the optimizer.
FSDP Warning: When using FSDP, several parameter groups will be conflated into a single one due to nested module wrapping and parameter flattening.

[lw3259111]

flash-attn v1 is work, but flash-attn v2 oom(same code)

[merrymercy]

Try to use transformers==4.28.1?