config_decoder.json

{
    "model_type": "sd3",
    "dynamo_backend": "inductor",
    "tunableop": false,
    "epochs": 10,
    "optimizer": "AdamW",
    "optimizer_args": {"weight_decay": 0.01, "eps": 1e-8, "betas": [0.9, 0.999], "fused": true},
    "lr_scheduler": "CosineAnnealingLR",
    "lr_scheduler_args": {"T_max": 102400, "eta_min": 4e-7},
    "learning_rate": 4e-6,
    "max_grad_norm": 0.0,
    "max_grad_clip": 0.0,
    "log_grad_stats": true,
    "batch_size": 1,
    "gradient_accumulation_steps": 16,
    "gradient_checkpointing": true,
    "weights_dtype": "float32",
    "mixed_precision": "fp16",
    "ema_decay": 0.995,
    "ema_update_steps": 32,
    "use_foreach_ema": true,
    "update_ema_on_cpu": false,
    "offload_ema_to_cpu": false,
    "offload_ema_non_blocking": true,
    "offload_ema_pin_memory": false,
    "ema_weights_dtype": "float32",
    "load_queue_lenght": 64,
    "max_load_workers": 8,
    "gc_steps": 256,
    "log_with": "wandb",
    "project_name": "anime_decoder",
    "resume_from": "none",
    "checkpoint_save_steps": 256,
    "checkpoints_limit": 20,
    "reshuffle": true,
    "dataset_index": "cache/decoder_dataset_index.json",
    "project_dir": "/mnt/DataSSD/AI/models/SoteDiffusion/StableDiffusion3/train_decoder",
    "model_path": "/mnt/DataSSD/AI/models/test/sd3/32_sb-3.6B_raifu_plus",
    "image_ext": ".jxl",
    "dataset_paths": [
        ["/mnt/DataSSD/AI/anime_image_dataset/vae/danbooru-best-png-latent", "/mnt/DataSSD/AI/anime_image_dataset/vae/danbooru-best-png-cropped", 1]
    ]
}