# @package _global_
# Shared offline/online inference preset. Set dit_weight here or override at launch.
defaults:
  - _self_

name: inference
output_dir: ./outputs/distill
dit_weight: /path/to/hy-nomas

model:
  dit:
    model_type: HY-Nomas-G/2-32324-MMDiT
    attn_mode: magi_mp_causal
    unpad_zero_tokens: true
    rope_theta: 256
    master_weight_type: bf16
  camera:
    use_camera_control: true
    injection_mode: extrinsics_attn
    per_segment_camera_attn: false
    per_chunk_camera_attn: true
    camera_chunk_size: 1
    strip_camera_text_blocks: true
  text_encoder:
    name: qwen-2.5vl-7b
    model_name: Qwen/Qwen2.5-VL-7B-Instruct
    tokenizer: qwen-2.5vl-7b
    text_states_dim: 3584
    text_len: 1280
    text_len_ll_cap: 1000
    text_projection: single_refiner
    prompt_template: li-dit-encode-image-json
    prompt_template_video: li-dit-encode-video-json
    hidden_state_skip_layer: 2
    use_attention_mask: true
    text_encoder_precision: bf16
  vae:
    name: pixelumm
    vae_precision: bf16
    vae_tiling: false
    vae_encode_tile_size: -1
    checkpoint: /path/to/pixelumm.safetensors
    parallel_mode: window
    output_rank: 0
    cpu_offload: false
    # Online demo enables stable-prefix decoding in its worker launcher.
    stable_prefix_streaming: false
    denoising_step_list: [0.999, 0.866, 0.634, 0.342]
    num_steps: 1
    seed: 42
    temporal_window_latents: 10
    temporal_overlap_latents: 5
    tile_pixels: 0
    tile_overlap_pixels: 64
    local_attention_tile_pixels: 128
    tail_hold_latents: 1
    deepgemm_fp8: false
    cutlass_fp8: false
    cute_attention: false
    compile_decoder: false

# These attention/prompt flags are shared by training and inference.
training:
  training_mode: inference_only
  cross_segment_text_head_only: true
  intra_segment_action_mask: true
  prev_seg_last_frame_only: false

dataset:
  json_prompt_normalize_action_timestamp: true
  json_prompt_clear_timestamps: false

parallelism:
  sp_size: 8

inference:
  prompt: null
  save_path: ./outputs/distill
  rollout_type: seg_ar
  inference_chunk_size: 0
  video_size: [720, 1280]
  video_fps: 24
  video_length: event
  infer_steps_video: 8
  infer_flow_shift_video: 5.0
  cfg_scale_video: 1.0
  camera_cfg_scale: 1.0
  solver_type: euler
  denoise_mode: ode
  seed: 42
  cache_at_step: -1
  cache_noise_sigma: 0.206897
  max_cache_segments: null
  ar_uncond_with_camera: true
  ar_single_slot_cfg_cache: true
  ar_cache_latent_cond_only: false
  use_global_camera_rope_offset: false
  seg_ar_segment_vae_decode: false
  max_segments: null
  max_iters: null
  skip_existing: true
  save_latents: true
  save_prompt_txt: true
  desc_field_first: ""
  desc_field_rest: ""
  no_desc_long_after_first: true
  global_only_in_first_segment: true
  resolve_subject_ids: false
  keep_keywords:
    - Style
    - perspective_view
    - subject_list
    - subject_list.subject_id
    - subject_list.display_name
    - subject_list.subject_attribute
    - segments
    - segments.visual_description_long
    - segments.description_change
    - segments.contained_actions
    - segments.contained_actions.start_time
    - segments.contained_actions.end_time
    - segments.contained_actions.action_description
  normalize_relative_to_first_frame: true
  normalize_framewise: true
  normalize_displacement: true
  trajectory_translation_scale: 1.0
  trajectory_rotation_scale: 2.5
  overlay_keys: true
  overlay_keys_keep_original: false
  overlay_keys_abs_threshold: 0.002
  neg_prompt: >-
    low quality, worst quality, low resolution, blurry, out of focus,
    jpeg artifacts, noise, grainy, distorted, deformed, bad anatomy,
    bad proportions, extra limbs, missing limbs, duplicate, watermark,
    text, logo, signature
  sparse_attention:
    enabled: false
    route_refresh_steps: [3, 6]  # Used only when Sparse Attention is enabled.
    dense_first_n_steps: 2
    dense_first_n_layers: 1
    global_keep_fraction: 0.25
    selection_scope: global_fraction
    router_pooling: 2x2
    router_mode: cute_bf16
    flex_sparse_layout: compact
    spatial_block_size: 8
    temporal_block_size: 1
    spatial_shift_mode: none
    query_time_group_size: 1
    force_self_block: true
    router_workspace_gib: 16
  sparse_optim:
    enabled: false