models:
  student:
    _target_: fastvideo.train.models.wan.WanModel
    init_from: /mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0361-14b-c256-ode500-260922-123321/export-step500
    trainable: true
    attention_backend: VSA_QAT_TRAIN_C256
    enable_gradient_checkpointing_type: full
    flow_shift: 8.0
  teacher:
    _target_: fastvideo.train.models.wan.WanModel
    init_from: Wan-AI/Wan2.1-T2V-14B-Diffusers
    trainable: false
    attention_backend: TORCH_SDPA
    disable_custom_init_weights: true
    flow_shift: 8.0
  critic:
    _target_: fastvideo.train.models.wan.WanModel
    init_from: Wan-AI/Wan2.1-T2V-14B-Diffusers
    trainable: true
    attention_backend: TORCH_SDPA
    disable_custom_init_weights: true
    enable_gradient_checkpointing_type: full
    flow_shift: 8.0
method:
  _target_: fastvideo.train.methods.distribution_matching.dmd2.DMD2Method
  rollout_mode: simulate
  generator_update_interval: 1
  real_score_guidance_scale: 6.0
  dmd_denoising_steps:
  - 1000.0
  - 941.1763916015625
  - 800.0
  min_timestep_ratio: 0.02
  max_timestep_ratio: 0.98
  fake_score_learning_rate: 4.0e-07
  fake_score_betas:
  - 0.0
  - 0.999
  fake_score_lr_scheduler: constant
  shifted_ode_schedule:
    version: 1
    teacher_steps: 48
    flow_shift: 8.0
    student_steps: 3
    indices:
    - 0
    - 16
    - 32
    timesteps:
    - 1000.0
    - 941.1763916015625
    - 800.0
    sigmas:
    - 1.0
    - 0.9411764144897461
    - 0.800000011920929
  warp_denoising_step: false
training:
  distributed:
    num_gpus: 16
    sp_size: 4
    tp_size: 1
    hsdp_replicate_dim: 1
    hsdp_shard_dim: 16
  data:
    data_path: /mnt/lustre/vlm-miz055/vsqa-14b/data/vidprom-14b-conditioning-full-canonical-r38ec498/training_dataset
    dataloader_num_workers: 0
    train_batch_size: 1
    training_cfg_rate: 0.0
    seed: 1000
    num_latent_t: 20
    num_height: 768
    num_width: 1280
    num_frames: 77
    preprocessed_data_type: text_only
  optimizer:
    learning_rate: 2.0e-06
    betas:
    - 0.0
    - 0.999
    weight_decay: 0.01
    lr_scheduler: constant
    lr_warmup_steps: 0
  loop:
    max_train_steps: 1000
    gradient_accumulation_steps: 4
  checkpoint:
    output_dir: /mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0362-14b-c256-dmd1000-260922-123321/checkpoints
    training_state_checkpointing_steps: 50
    checkpoints_total_limit: 0
    resume_from_checkpoint: ''
  tracker:
    trackers:
    - wandb
    project_name: fastvideo
    run_name: fastvideo-v0362-14b-c256-dmd1000-260922-123321
    entity: alexzms-ucsd
  vsa:
    sparsity: 0.9
    cache_tile_buf: false
  dit_precision: fp32
callbacks:
  grad_clip:
    max_grad_norm: 1.0
  validation:
    pipeline_target: fastvideo.pipelines.basic.wan.wan_shifted_dmd_pipeline.WanShiftedDMDPipeline
    dataset_file: /mnt/lustre/vlm-miz055/vsqa-14b/runs/fastvideo-v0362-14b-c256-dmd1000-260922-123321/config/validation_16.json
    every_steps: 50
    run_at_start: false
    sampling_steps:
    - 3
    guidance_scale: 1.0
    height: 768
    width: 1280
    num_frames: 77
    attn_qat_infer: false
    _target_: fastvideo.train.methods.ode_init.dmd_validation.ShiftedDMDValidationCallback
    use_validation_media_conditioning: false
    offload_training_state: true
    unload_pipeline_after_validation: true
pipeline:
  flow_shift: 8.0
  dit_config:
    quant_config: nvfp4_qat_train
  dmd_denoising_steps:
  - 1000.0
  - 941.1763916015625
  - 800.0
  warp_denoising_step: false