Question: Using the LLaVA-NEXT Dataset for SFT

[nihaotian1]

Hi,
thanks for your great work.

While attempting to reproduce the LLaDA-V ablation study results, I successfully executed scripts/llada_v_pretrain.sh. However, I then encountered several issues in the subsequent steps.

1. I attempted to increase the per_device_batch_size from 1 to 8 while running bash scripts/train_ablation/llada_v_sft.sh. Although GPU memory usage increased, the total training time remained unchanged at approximately 72 hours with 4 GPUs.

2. I observed that the grad_norm was reported as 0.0 during training. Is this expected, or could it indicate a problem? (The only modifications I made were to the training scripts; the rest of the original code remains unchanged.)

[yyyouy]

hello, have you resolved this question? sorry for my late reply