{
  "_class_name": "sprint_dual",
  "model_type": "sprint_dual",
  "in_channels": 32,
  "hidden_size": 768,
  "depth": 13,
  "num_heads": 12,
  "encoder_depth": 2,
  "decoder_depth": 2,
  "drop_ratio": 0.0,
  "drop_target": "image",
  "residual_type": "concat_linear",
  "hf_text_encoder": "LiquidAI/LFM2.5-230M",
  "hf_vae": "kaiyuyue/FLUX.2-dev-vae",
  "vae_mean": 0.05106564902828182,
  "vae_std": 1.69915383699,
  "resume_from_checkpoint": "epoch_99_step_132000/epoch_99_step_132000",
  "cfg_mask_prob": 0.0,
  "skip_checkpointing_layers": 8,
  "use_random_drop": false,
  "vae_batch_size": 16,
  "use_calibrated_spatial": true,
  "train": {
    "batch_size": 64,
    "lr": 0.0005,
    "epochs": 150,
    "gradient_accumulation_steps": 4,
    "dtype": "fp32",
    "use_checkpointing": true,
    "train_te": false,
    "wd": 0.01,
    "warmup": 0.04,
    "min_lr_ratio": 0.02,
    "reset_scheduler": false,
    "reset_optimizer": false,
    "use_cos_scheduler": false,
    "objective": "flow_matching",
    "shift": 1.0,
    "snr_gamma": 5.0,
    "input_perturb": 0.1,
    "cfg_dropout_prob": 0.1,
    "seed": 232,
    "use_bitsandbytes": true,
    "train_only_output": false,
    "skip_save_n_times": 4,
    "run_stability_check": false,
    "world_size": 1,
    "use_ot": false,
    "use_ema": false,
    "ema_decay": 0.999,
    "use_kahan_sum": false,
    "use_cpu_accumulator": false,
    "compile_model": true,
    "gpu_peak_tflops": 165.2
  },
  "data": {
    "dataset_type": "streaming",
    "h5_path": "data",
    "streaming_dataset_name": "aipracticecafe-mirror/curated-danbooru-2026-256px-flux2-vae",
    "num_workers": 6,
    "prefetch_factor": 6,
    "load_into_ram": true,
    "tag_dropout": 0.1,
    "use_shift_crop": true,
    "resolution": 256,
    "buffer_size": 5000,
    "cache_latents_to_ram": true,
    "is_latent": true,
    "precompute_num_workers": 2,
    "precompute_prefetch_factor": 2
  },
  "models": {
    "model_type": "sprint_dual",
    "resume_from_checkpoint": "epoch_99_step_132000/epoch_99_step_132000",
    "output_head_path": null,
    "hf_text_encoder": "LiquidAI/LFM2.5-230M",
    "hidden_size": 768,
    "depth": 13,
    "num_heads": 12,
    "encoder_depth": 2,
    "decoder_depth": 2,
    "drop_ratio": 0.0,
    "drop_target": "image",
    "residual_type": "concat_linear",
    "cfg_mask_prob": 0.0,
    "skip_checkpointing_layers": 8,
    "use_random_drop": false,
    "hf_vae": "kaiyuyue/FLUX.2-dev-vae",
    "vae_mean": 0.05106564902828182,
    "vae_std": 1.69915383699,
    "vae_batch_size": 16,
    "in_channels": 32,
    "use_calibrated_spatial": true
  },
  "paths": {
    "models": "models"
  },
  "logging": {
    "project": "dual-dit",
    "wandb_entity": "aipracticecafe-aipracticecafe",
    "save_interval": 13000,
    "hf_repo": "dual-stream-dit",
    "wandb_run_name_prefix": "4090-sprint-latents-spatial-fine"
  },
  "sampling": {
    "interval": 1300,
    "config_file": "sample.toml"
  }
}