config.yaml

# Quantization configuration for Wan2.2 I2V model
# W4A4 NVFP4: 4-bit FP4 (E2M1) weights and activations using NVFP4 format
#
# Full Q-VDiT approach:
# - PTQ: Quantize ALL Linear layers with NVFP4 weights (per-channel) and activations (per-token)
# - TQE: Token-aware Quantization Estimator (LoRA-like correction, already in QuantLayer)
# - TMD: Temporal Maintenance Distillation (KL divergence on frame similarity)

# Layers to keep in full precision (input/output projections + first/last blocks)
part_fp_list: "./t2v/configs/quant/wan/remain_fp.txt"

# Model identification
model:
  model_id: "wan_i2v_14b"
  model_type: 'wan'

# Conditional generation flag
conditional: True

# Calibration data settings (reduced batch_size for Wan 14B due to memory constraints)
calib_data:
  path: null  # Set via command line
  n_steps: 10  # Number of timesteps to use for calibration
  batch_size: 1  # Reduced from 4 for Wan 14B model (saves ~60GB GPU memory)
  n_samples: 3   # Number of samples per timestep

# Quantization settings
quant:
  # Weight quantization (NVFP4 E2M1 per-channel)
  weight:
    quantizer:
      quant_type: 'nvfp4'  # Use NVFP4 E2M1 floating-point quantizer
      n_bits: 4
      per_group: 'channel'  # Per-channel quantization for weights
      scale_method: 'absmax'
    optimization:
      iters: 1000  # Increased from 200 - sufficient iterations for LoRA/delta convergence
      use_grad: False
      loss:
        # TMD (Temporal Maintenance Distillation) - uses frame-wise similarity preservation
        # Works with Wan's 5D latent tensors [B,C,T,H,W]
        reconstruction_loss_type: 'relation'
        lambda_coeff: 1.0
        b_range: [10, 2]
        warmup: 0.0
        decay_start: 0.0
        p: 2.0
      params:
        delta:
          lr: 1.e-6  # Q-VDiT: 1e-6 for weight scale params

  # Activation quantization (NVFP4 E2M1 dynamic per-token)
  activation:
    quantizer:
      quant_type: 'nvfp4'  # Use NVFP4 E2M1 floating-point quantizer
      n_bits: 4
      per_group: 'token'  # Per-token quantization for activations
      dynamic: True  # Dynamic quantization (compute scale on-the-fly with STE)
      scale_method: 'absmax'
      # Token count configuration for Wan model
      n_tokens: 5120  # Combined spatial-temporal tokens
      n_text_tokens: 512  # Text encoder sequence length
      n_image_tokens: 257  # CLIP image encoder (256 patches + 1 CLS)
      # Smooth quantization settings
      # NOTE: Using single timerange is more robust for Wan2.2 since each transformer
      # only sees specific timestep ranges. Multiple timeranges risk uncovered ranges.
      smooth_quant:
        enable: True
        channel_wise_scale_type: 'momentum_act_max'
        momentum: 0.95
        alpha: [0.11]  # Single alpha for single timerange
        timerange: [[0, 1000]]  # Single timerange - simpler and more robust

  # TQE (Token-aware Quantization Estimator) parameters
  # LoRA-like low-rank correction already implemented in QuantLayer
  tqe:
    lr: 1.e-5  # Q-VDiT: 1e-5 for TQE (LoRA) params
    # Memory optimization: process layers in batches during M initialization
    layer_batch_size: 50  # Number of layers to process at once (reduces peak memory)
    # Optional: filter layers for TQE M initialization
    # 'attention' = only init M for attention layers (.attn1., .attn2.)
    # null = init M for all QuantLayers (default)
    layer_filter: 'attention'  # Reduces from ~1200 layers to ~240 layers

  # Gradient checkpointing (recommended for 14B model)
  grad_checkpoint: True

# Timestep-wise quantization (optional)
timestep_wise: False

# CFG (Classifier-Free Guidance) split handling
cfg_split: False