[Usage] After fine-tuning LLaVA 1.5, mm_projector.bin file is not available

[rayluo88]

Describe the issue

Issue: I have fine tuned the llava-v1.5-7b with lora. And in the output directory I got some files.

• adapter_model.safetensors
• config.json
• README.md
• adapter_config.json
• non_lora_trainables.bin
• trainer_state.json

The mm_projector.bin file is missing, which is required by scripts/merge_lora_weights.py and run_llava.py (llava/eval/run_llava.py)

The mm_projector.bin file contains the projector weights, right? How to generate/extract this file? @haotian-liu

Command:

deepspeed llava/train/train_mem.py \
    --lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 2e-5 \
    --deepspeed $DEEPSPEED_JSON \
    --model_name_or_path $MODEL_NAME \
    --version v1 \
    --data_path $DATA_PATH \
    --image_folder $IMAGE_FOLDER \
    --vision_tower $VISION_TOWER \
    --mm_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --image_aspect_ratio pad \
    --group_by_modality_length True \
    --bf16 True \
    --output_dir $OUTPUT_DIR \
    --num_train_epochs 1 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --save_strategy "steps" \
    --save_steps 50000 \
    --save_total_limit 1 \
    --learning_rate 2e-4 \
    --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 wandb

Log:

PASTE THE LOGS HERE.

Screenshots:
You may attach screenshots if it better explains the issue.

[DemonsAH]

Add 'lora' in the folder name helped mine.
Merged model could run inference and showed the affection of fine-tuning.
I assume you can run inference without merging if you add 'lora' in the folder name.

[HuizhaoWang]

Describe the issue

Issue: I have fine tuned the llava-v1.5-7b with lora. And in the output directory I got some files.

• adapter_model.safetensors
• config.json
• README.md
• adapter_config.json
• non_lora_trainables.bin
• trainer_state.json

The mm_projector.bin file is missing, which is required by scripts/merge_lora_weights.py and run_llava.py (llava/eval/run_llava.py)

The mm_projector.bin file contains the projector weights, right? How to generate/extract this file? @haotian-liu

Command:

deepspeed llava/train/train_mem.py \
    --lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 2e-5 \
    --deepspeed $DEEPSPEED_JSON \
    --model_name_or_path $MODEL_NAME \
    --version v1 \
    --data_path $DATA_PATH \
    --image_folder $IMAGE_FOLDER \
    --vision_tower $VISION_TOWER \
    --mm_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --image_aspect_ratio pad \
    --group_by_modality_length True \
    --bf16 True \
    --output_dir $OUTPUT_DIR \
    --num_train_epochs 1 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --save_strategy "steps" \
    --save_steps 50000 \
    --save_total_limit 1 \
    --learning_rate 2e-4 \
    --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 wandb

Log:

PASTE THE LOGS HERE.

Screenshots: You may attach screenshots if it better explains the issue.

I also encountered this problem, have you solved it yet?