[Question] add extral layers in llava

[NONGFUYULANG]

Question

I have tried to add more linear layer to deal with some extra features. And i only want to train this layers. and i uses the scripts below to finetune the whole model, but when i tried to print the grad and weight. i found that grad is none and weight does not change. So i want to figure out that: is this the problem of lora or something else. And i have alread set the require_grads parameter to True in the training.py file.

scripts

deepspeed train.py \ --lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 2e-5 \ --deepspeed /mnt/modified_llavas/LLaVA/scripts/zero3.json \ --model_name_or_path liuhaotian/llava-v1.5-7b \ --version v2 \ --data_path /mnt/f/modified_llavas/LLaVA/llava_out_new.json \ --image_folder test \ --vision_tower openai/clip-vit-large-patch14-336 \ --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 /mnt/f/modified_llavas/LLaVA/llava_fullweights \ --num_train_epochs 1 \ --per_device_train_batch_size 1 \ --per_device_eval_batch_size 1 \ --gradient_accumulation_steps 1 \ --evaluation_strategy "no" \ --save_strategy "steps" \ --save_steps 5 \ --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 4096 \ --gradient_checkpointing True \ --dataloader_num_workers 4 \ --lazy_preprocess True \ --report_to wandb

add extra layers

and i have used the test_head layer in the llava_arch.py file. Does anyone have ideas about it?

[zihui-debug]

can you provide the forward code and other sections modified?

[WayneTomas]

The grad is None is because the deepspeed split the model to different cards. If you want to obtain the grad, you should use the functions provided in deepspeed it-self to colloct the grad, and the traditional "tensor.grad" is not working in deepspeed.