Why kohya_ss faster compare to sd-scripts repository?

Hi everyone, i have tried to run both the kohya_ss webui and script in sd-scripts repo with below setting:
`"sdxl": false, "v2": false, "v_parameterization": false, "num_cpu_threads_per_process": 4, "pretrained_model_name_or_path": "deliberate_v3.safetensors", "color_aug": false, "flip_aug": false, "random_crop": false, "resolution": "512,512", "cache_latents": true, "cache_latents_to_disk": false, "enable_bucket": true, "min_bucket_reso": 256, "max_bucket_reso": 2048, "bucket_reso_steps": 64, "bucket_no_upscale": true, "reg_data_dir": "", "dataset_repeats": 20, "save_precision": "bf16", "save_every_n_epochs": 1, "save_last_n_epochs": 1, "save_state": false, "train_batch_size": 10, "max_token_length": 150, "xformers": true, "vae": "", "max_train_epochs": 7, "seed": 1, "gradient_checkpointing": true, "gradient_accumulation_steps": 1, "mixed_precision": "bf16", "clip_skip": 1, "noise_offset": 0, "multires_noise_iterations": 0, "multires_noise_discount": 0, "adaptive_noise_scale": 0, "min_timestep": 0, "max_timestep": 1000, "prior_loss_weight": 1.0, "optimizer_type": "AdamW8bit", "learning_rate": 0.0001, "lr_scheduler": "cosine", "lr_warmup_steps": 10, "min_snr_gamma": 0, "scale_v_pred_loss_like_noise_pred": false, "v_pred_like_loss": 0, "weighted_captions": false, "save_model_as": "safetensors", "unet_lr": 0.0001, "text_encoder_lr": 5e-05, "network_module": "networks.lora", "network_dim": 256, "network_alpha": 256, "network_dropout": 0, "network_train_unet_only": false, "network_train_text_encoder_only": false, "dim_from_weights": false, "scale_weight_norms": 0, "no_half_vae": false`
I have also set accelerate config to be the same as config_files/accelerate/default_config.yaml. I want to ask why the kohya-ss is much faster compare to sd-scripts repository? It took 5 minutes in kohya-ss and ~10 minutes in sd-scripts.

Compare the kohya_ss webui and the sd-scripts repository using the settings in the report, including config_files/accelerate/default_config.yaml. First reproduce the timing difference and collect the missing environment and version details; done means identifying an evidence-based cause for the roughly five-minute versus ten-minute runtime difference.

Written by the indexing model from the issue text.