When will the training code open sourced?

[jimchenhub]

When will the training code open sourced?

[Luodian]

We definitely have the plan to opensource everything (including previous LLaVA-NeXT's training code, data) to benefit community.

However, there're still lot of works to do and more releases to be expected~

[JianbangZ]

@jimchenhub @Luodian
I found the training scripts was publiced in the HF model README
https://huggingface.co/lmms-lab/llama3-llava-next-8b

Any way to actually run this training script?

[jimchenhub]

Can you please tell me if there is any obvious difference between the training code and the one in https://github.com/haotian-liu/LLaVA? I'm trying to finetune it on my data set.

[Luodian]

There's no major difference, if you wish to finetune our 8b/72b/110b. Only diff is that you need to apply new conversation templates.

[jimchenhub]

Indeed, I found that the special tokens of llama3-8B are different from those of vicuna-7B. Could you please tell me what the conversation template is like you use to finetune llama3-8B?

[Luodian]

you can find in llava/conversations.py, we use llama-3.

But it's for inference, in training, you need to implement the masking logic for llama-3 your side.

[jimchenhub]

OK thanks, I'll try it.

[jimchenhub]

Hi,

From your provided checkpoint (https://huggingface.co/lmms-lab/llama3-llava-next-8b), I found that the pre-trained config is

PROMPT_VERSION=plain
PRETRAIN_DATA_VERSION="blip558k"

So I referred to https://github.com/haotian-liu/LLaVA/blob/main/scripts/v1_5/pretrain.sh to pre-train with LLaMA3-8B backend as the following script

deepspeed llava/train/train_mem.py \
    --deepspeed ./scripts/zero2.json \
    --model_name_or_path ckpts/Meta-Llama-3-8B \
    --version plain \
    --data_path ./playground/data/LLaVA-Pretrain/blip_laion_cc_sbu_558k.json \
    --image_folder ./playground/data/LLaVA-Pretrain/images \
    --vision_tower ckpts/clip-vit-large-patch14-336 \
    --mm_projector_type mlp2x_gelu \
    --tune_mm_mlp_adapter True \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --bf16 True \
    --output_dir ./checkpoints/llava-v1.5-llama-8b-pretrain \
    --num_train_epochs 1 \
    --per_device_train_batch_size 32 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --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 tensorboard

However, the training loss is much larger than pre-trained with vicuna-7B (2.40 v.s. 2.06).

To adapt LLaMA3-8B to the training code in https://github.com/haotian-liu/LLaVA, I manually add pad_token and unk_token like

if tokenizer.pad_token is None:
        print("\n add unk_token \n\n")
        smart_tokenizer_and_embedding_resize(
            special_tokens_dict=dict(
                unk_token="<unk>"
            ),
            tokenizer=tokenizer,
            model=model,
        )
tokenizer.pad_token = tokenizer.unk_token

I wonder whether it will cause bad pre-train performance or not? If this is not correct, how can I set padding_value in this line input_ids = torch.nn.utils.rnn.pad_sequence()?

[zihaolucky]

Hi @jimchenhub have you figured it out?