model_type: NAVA
modality: audio_video # [audio, video, image, audio_video]

# 模型
model_id: /root/paddlejob/workspace/Qwen3-1.7B
# model_id: GPT-B
use_bf16: true
audio_latent_ch: 128
video_latent_ch: 48

patch_size: 2
image_size: 960
log_width: 1280
log_height: 704
lambda_ddpm: 1.0

pipeline: nava_src.pipeline_nava.AudioVideoPipeline
scheduler_personalized: true
scheduler_unipc: true
use_loss_reweight: true
attention_only: false

use_mmdit_model: true
audio_loss_coff: 0.2
vision_loss_coff: 1

# 推理
video_guidance_scale: 3.0
audio_guidance_scale: 2.0
video_align_guidance_scale: 3.0
audio_align_guidance_scale: 2.0
masking_modality_prob: 0.2
align_3d_cfg: true
i2v_mode_prob: 0.5
no_split_norm_ffn: true
timbre_cfg: true
timbre_align_guidance_scale: 3.0

# 数据
data:
  # data_filelist: data/video_text/data_videogen_0630_256px_shijue.list
  # data_weights: data/video_text/data_video_0630_shijue.weight
  # data_filelist: data/audio_text/mixture_0207/data.list
  # data_weights: data/audio_text/mixture_0207/data.weight
  # data_filelist: av_data_v2.list
  # data_weights: av_data.weight
  data_filelist: av_data_v4_470w_filter10_with_audio.list
  data_weights: av_data_v4_470w_filter10_with_audio.weight
  # data_filelist: joint_train_mmdit.list
  # data_weights: joint_train_mmdit.weight
  use_length_buckets: true
  num_length_buckets: 5
  enable_ddp_bucket_sync: true
  audio_tokens_per_sec: 25

  queue_size: 16
  io_workers: 4

  modal_prob: 
    text_to_audio: 1
    text_to_video: 0.0
    text_to_image: 0.0
    text_to_av: 1

  # server config
  use_server: False
  audio_scaling_factor: 1.0
  audio_shift_factor: 0.0
  video_scaling_factor: 1.0
  video_shift_factor: 0.0
  zlib_mode: false
  zstd_mode: true

  # video_fps: 16
  # video_min_frames: 33
  # video_max_frames: 300
  # video_tgt_frames: 81
  min_audio_duration: 0.0
  max_audio_duration: 10.0

  video_fps: 24
  video_min_frames: 33
  video_max_frames: 400
  video_tgt_frames: 121
  # video_tgt_frames: 241

  add_spk_emb: true
  spk_emb_prob: 0.9
  use_speech_special_token: false

model:
  shift: 5
  shift_audio: 5
  num_train_timesteps: 1000
  gradient_checkpointing: true
  cross_gate: false
  cross_1d_rope: false

  # audio_config: nava_src/models/nava/configs/model/dit/audio.json
  # video_config: nava_src/models/nava/configs/model/dit/video.json
  audio_config: nava_src/models/nava/configs/model/dit/audio_tiny.json
  video_config: nava_src/models/nava/configs/model/dit/video_tiny.json
  joint_config: nava_src/models/nava/configs/model/dit/NAVA_6B.json
  ckpt_dir: ./
  audio_vae_ckpt_dir: ./huggingface_upload/params


# 训练
batch_size: 1
lr: 1.0e-4
weight_decay: 0
warmup_steps: 0
max_steps: 1000000
save_every: 2500
out_dir: outputs/t2av_baseline_demo_mmdit_test_no_separate_modulation
seed: 42

flow: true

# 性能相关
num_workers: 0
prefetch_factor: 32
grad_accum_steps: 1
log_every: 1
log_cases_every: 1000000        # >0 才会写图像
log_sample_steps: 25
log_cfg: 4
max_grad_norm: 1.0
amp_dtype: bf16            # bf16 / fp16 / null