# Wan Models Training - Optimal Defaults Configuration
# This file contains the recommended default values for Wan model training
# Supports all Wan models: T2V, I2V, T2I, FLF2V, Fun-Control, and Wan 2.2 Advanced models
# Based on Wan documentation and best practices

# Training Mode Selection
training_mode = "LoRA Training"  # The installed Wan backend supports LoRA/network training.

# Wan Model Selection - REQUIRED
task = "t2v-14B"  # REQUIRED: Wan model type (t2v-14B, t2v-1.3B, i2v-14B, t2i-14B, flf2v-14B, t2v-1.3B-FC, t2v-14B-FC, i2v-14B-FC, t2v-A14B, i2v-A14B)

# Model Settings - REQUIRED FIELDS MUST BE SET BY USER
dit = ""  # REQUIRED: Path to DiT checkpoint (.safetensors)
vae = ""  # REQUIRED: Path to VAE checkpoint (Wan2.1_VAE.pth or wan_2.1_vae.safetensors)
t5 = ""  # REQUIRED: Path to T5 text encoder (umt5-xxl-enc-bf16.safetensors)
clip = ""  # REQUIRED ONLY for i2v-14B, flf2v-14B, i2v-14B-FC models. Path to CLIP vision encoder (models_clip_open-clip-xlm-roberta-large-vit-huge-14.safetensors)
dataset_config = ""  # Path to dataset TOML configuration (required if using "Use TOML File" mode)
dataset_config_mode = "Generate from Folder Structure"  # Default to auto-generate from folder structure

# Wan 2.2 Advanced Models (A14B) - Dual Model System
dit_high_noise = ""  # [WAN 2.2 ONLY] Path to high noise DiT model for Wan 2.2 Advanced models (t2v-A14B, i2v-A14B)
timestep_boundary = 0.0  # [WAN 2.2 ONLY] Timestep boundary for dual model system. Auto-detected from model config if not set
offload_inactive_dit = false  # [WAN 2.2 ONLY] Offload inactive DiT to CPU to save VRAM during dual model training
use_pinned_memory_for_block_swap = false  # Use Pinned Memory for Block Swapping: Uses more system RAM but speeds up training. The speed up maybe significant depending on system settings. To work, go to Advanced Graphics settings in System > Display > Graphics as in tutorial video and disable Hardware-Accelerated GPU Scheduling and restart your PC. Only effective when blocks_to_swap > 0
block_swap_h2d_only = false  # LoRA only; requires blocks_to_swap > 0 and gradient_checkpointing; use standard swap for Wan 2.2 dual-DiT
block_swap_ring_size = 2  # 2 overlaps transfer/compute; 1 minimizes VRAM

# Folder Structure Settings (used when dataset_config_mode = "Generate from Folder Structure")
parent_folder_path = ""  # REQUIRED when using folder structure mode: Path to parent folder containing training images/videos
dataset_resolution_width = 960  # Target width for training. Optimal resolutions: 960×960, 1280×720, 720×1280
dataset_resolution_height = 960  # Target height for training. Optimal resolutions: 960×960, 1280×720, 720×1280
dataset_caption_extension = ".txt"  # Extension for caption files
create_missing_captions = true  # Auto-create empty captions for images without caption files
caption_strategy = "folder_name"  # How to handle missing captions: "folder_name" or "empty"
dataset_batch_size = 1  # Batch size for dataset processing
dataset_enable_bucket = false  # Enable bucketing for variable aspect ratios
dataset_bucket_no_upscale = false  # Prevent upscaling in bucket mode
dataset_cache_directory = "cache_dir"  # Directory for caching processed data
generated_toml_path = ""  # Path where generated TOML will be saved (auto-set)

# Video-specific Settings (for T2V, I2V, FLF2V models)
num_frames = 81  # Number of frames for video training (81 is default for Wan models)
one_frame = false  # Enable one-frame training mode for image-like training on video models

# Video Frame Extraction Settings
frame_extraction = "head"  # Frame extraction method: "head" (recommended), "chunk", "slide", "uniform", "full"
frame_stride = 1  # Step size for sliding window extraction (used with "slide" method)
frame_sample = 1  # Number of samples to extract (used with "uniform" method)
# Dataset generation target_frames (written into generated dataset TOML for video datasets)
target_frames = 81  # Target video frame count for dataset generation (saved as target_frames=[N]). Must be N*4+1 (1,5,9,...,81)
auto_normalize_target_frames = true  # Scan dataset and clamp Target Frames to the minimum video length found (then round to N*4+1)
max_frames = 129  # Maximum number of frames to extract from any video (used with "full" method)
source_fps = 0  # Original video FPS for frame rate conversion (0=auto-detect)

# Data Types and Precision
dit_dtype = "bfloat16"  # DiT model data type. bfloat16=best quality, float16=faster
text_encoder_dtype = "bfloat16"  # T5 text encoder data type
vae_dtype = "bfloat16"  # VAE data type
clip_vision_dtype = "bfloat16"  # CLIP vision encoder data type

# Memory Optimization Settings
fp8_base = false  # Enable FP8 for DiT model to reduce VRAM usage (requires fp8_scaled=true)
fp8_scaled = false  # REQUIRED when fp8_base=true, provides better quality than standard FP8
fp8_t5 = false  # Enable FP8 for the T5 text encoder to reduce VRAM usage
blocks_to_swap = 0  # 0=disabled. Max varies by model. Higher values save more VRAM but require more RAM

# VAE Optimization Settings
vae_tiling = false  # Enable spatial tiling to reduce VRAM usage during VAE operations
vae_chunk_size = 0  # 0=auto/disabled. Higher=faster but more VRAM
vae_spatial_tile_sample_min_size = 0  # 0=disabled. 256=typical. Auto-enables vae_tiling if set

# Flow Matching Settings (Wan uses Flow Matching instead of diffusion)
timestep_sampling = "uniform"  # Timestep sampling method: "uniform", "sigmoid", "shift"
discrete_flow_shift = 1.0  # Discrete flow shift parameter
weighting_scheme = "none"  # Loss weighting scheme: "none", "mode"
logit_mean = 0.0  # Logit mean for timestep sampling
logit_std = 1.0  # Logit standard deviation for timestep sampling
mode_scale = 1.29  # Mode scale for weighting scheme

# Advanced Timestep Parameters
sigmoid_scale = 1.0  # Scale factor for sigmoid timestep sampling
min_timestep = 0  # Minimum timestep constraint (0=no constraint)
max_timestep = 1000  # Maximum timestep constraint (1000=no constraint)
preserve_distribution_shape = false  # Preserve original distribution when using min/max constraints
num_timestep_buckets = 0  # 0=disabled. 4-10=bucketed sampling for uniform distribution

# Training Settings
sdpa = true  # Use PyTorch's scaled dot product attention (recommended)
flash_attn = false  # Use Flash Attention (requires installation)
sage_attn = false  # Use Sage Attention (inference only)
xformers = false  # Use xFormers attention
split_attn = false  # Split attention computation to reduce peak memory use
use_legacy_sdpa = false  # Force the older PyTorch SDPA path instead of automatic verified acceleration
max_train_steps = 90000  # Maximum training steps
max_train_epochs = 200  # Maximum training epochs
max_data_loader_n_workers = 2  # Number of data loader workers
persistent_data_loader_workers = true  # Keep data loader workers persistent
seed = 99  # Random seed for reproducibility
gradient_checkpointing = true  # Enable gradient checkpointing to save VRAM
gradient_accumulation_steps = 1  # Number of gradient accumulation steps
full_bf16 = false  # EXPERIMENTAL: Store gradients in BF16 to save VRAM (currently disabled in backend)
full_fp16 = false  # EXPERIMENTAL: Store gradients in FP16 to save VRAM (currently disabled in backend)

# Optimizer Settings
optimizer_type = "adamw8bit"  # Optimizer type (adamw8bit recommended for memory efficiency)
optimizer_args = []  # Additional optimizer arguments
learning_rate = 1e-4  # Learning rate (1e-4 is good default for Wan models)
max_grad_norm = 1.0  # Maximum gradient norm for clipping
lr_scheduler = "constant"  # Learning rate scheduler type
lr_warmup_steps = 0  # Number of warmup steps (0=no warmup)
lr_decay_steps = 0  # Number of decay steps (0=no decay)
lr_scheduler_num_cycles = 1  # Number of scheduler cycles
lr_scheduler_power = 1.0  # Scheduler power parameter
lr_scheduler_timescale = 0  # Scheduler timescale (0=auto)
lr_scheduler_min_lr_ratio = 0.0  # Minimum learning rate ratio
lr_scheduler_type = ""  # Additional scheduler type
lr_scheduler_args = []  # Additional scheduler arguments

# Network Settings (LoRA Mode Only - Ignored in DreamBooth Mode)
no_metadata = false  # Disable metadata saving
network_weights = ""  # Path to pretrained LoRA weights to continue training
network_module = "networks.lora_wan"  # Network module for LoRA (auto-set based on training mode)
network_dim = 16  # LoRA network dimension/rank (16 recommended for Wan models)
network_alpha = 16.0  # LoRA alpha parameter (typically equal to network_dim)
network_dropout = 0.0  # LoRA dropout rate (0.0=no dropout)
network_args = []  # Additional network arguments
training_comment = ""  # Training comment for metadata
dim_from_weights = false  # Extract dimensions from existing weights
scale_weight_norms = 0.0  # Scale weight norms (0.0=disabled)
base_weights = ""  # Path to base LoRA weights to merge
base_weights_multiplier = 1.0  # Multiplier for base weights

# Save/Load Settings
output_dir = ""  # Output directory for saved models
output_name = "my-wan-lora"  # Base filename for saved models
resume = ""  # Path to checkpoint to resume training from
save_precision = "bf16"  # LoRA output dtype; BF16 keeps files about half the size of FP32
save_every_n_epochs = 10  # Save checkpoint every N epochs
save_every_n_steps = 0  # Save checkpoint every N steps (0=disabled)
save_last_n_epochs = 0  # Keep only last N epoch checkpoints (0=keep all)
save_last_n_epochs_state = 0  # Keep only last N epoch states (0=keep all)
save_last_n_steps = 0  # Keep only last N step checkpoints (0=keep all)
save_last_n_steps_state = 0  # Keep only last N step states (0=keep all)
save_state = false  # Save training state
save_state_on_train_end = false  # Save state when training ends
mem_eff_save = false  # Memory efficient saving for DreamBooth fine-tuning mode

# Caching Settings - Latents
caching_latent_device = "cuda"  # Device for latent caching
caching_latent_batch_size = 4  # Batch size for latent caching
caching_latent_num_workers = 8  # Number of workers for latent caching
caching_latent_skip_existing = true  # Skip existing cached latents
caching_latent_keep_cache = true  # Keep cache after training
caching_latent_debug_mode = ""  # Debug mode for latent caching
caching_latent_console_width = 80  # Console width for debug output
caching_latent_console_back = ""  # Console background for debug
caching_latent_console_num_images = 0  # Number of images in debug console (0=no limit)

# Caching Settings - Text Encoder
caching_teo_text_encoder1 = ""  # Optional T5 override for caching; empty uses the model T5 path
caching_teo_text_encoder2 = ""  # Reserved optional second text encoder path
caching_teo_device = "cuda"  # Device for text encoder caching
caching_teo_fp8_llm = false  # Use FP8 for T5 text encoder caching
caching_teo_batch_size = 16  # Batch size for text encoder caching
caching_teo_num_workers = 8  # Number of workers for text encoder caching
caching_teo_skip_existing = true  # Skip existing cached outputs
caching_teo_keep_cache = true  # Keep cache after training

# Torch compile (optional)
compile = false
compile_backend = "inductor"
compile_mode = "default"
compile_dynamic = "auto"  # "auto" | "true" | "false"
compile_fullgraph = false
compile_cache_size_limit = 0  # 0 = use PyTorch default

# Accelerate Launch Settings
mixed_precision = "bf16"  # Mixed precision training (bf16 recommended for Wan)
multi_gpu = false  # Enable multi-GPU training
gpu_ids = "0"  # GPU IDs for distributed training
num_processes = 1  # Number of processes for distributed training
num_machines = 1  # Number of machines for distributed training
num_cpu_threads_per_process = 2  # CPU threads per process
main_process_port = 0  # Port for distributed communication (0=auto)
dynamo_backend = "no"  # PyTorch dynamo backend (no=disabled)
dynamo_mode = ""  # Dynamo optimization mode
dynamo_use_fullgraph = false  # Use fullgraph mode for dynamo
dynamo_use_dynamic = false  # Use dynamic mode for dynamo
extra_accelerate_launch_args = ""  # Additional accelerate launch arguments

# Logging Settings
logging_dir = ""  # Directory for training logs
log_prefix = ""  # Prefix for log files
log_tracker_name = ""  # Name for experiment tracking
wandb_run_name = ""  # Weights & Biases run name
log_tracker_config = ""  # Configuration for experiment tracking
wandb_api_key = ""  # Weights & Biases API key
log_config = false  # Log configuration to tracker

# DDP Settings (Distributed Training)
ddp_timeout = 0  # DDP timeout in minutes (0=default)
ddp_gradient_as_bucket_view = false  # Use gradient as bucket view
ddp_static_graph = false  # Use static graph for DDP

# Sample Generation Settings
sample_every_n_steps = 0  # Generate samples every N steps (0=disabled)
sample_every_n_epochs = 0  # Generate samples every N epochs (0=disabled)
sample_at_first = false  # Generate samples before training starts
sample_prompts = ""  # Path to file with sample prompts
disable_prompt_enhancement = false  # true = use prompt file exactly as written (no auto-added defaults)

# Default Sample Parameters for Video Generation
sample_width = 960  # Sample video width. Optimal resolutions: 960×960, 1280×720, 720×1280
sample_height = 960  # Sample video height. Optimal resolutions: 960×960, 1280×720, 720×1280
sample_num_frames = 81  # Number of frames in sample videos
sample_steps = 20  # Number of inference steps for samples
sample_guidance_scale = 7.0  # Default WAN training sample CFG scale (written as --l in enhanced prompts)
sample_seed = 99  # Seed for sample generation (-1=random)
sample_negative_prompt = ""  # Default negative prompt for samples

# Metadata Settings
metadata_author = ""  # Author metadata
metadata_description = ""  # Description metadata
metadata_license = ""  # License metadata
metadata_tags = ""  # Tags metadata
metadata_title = ""  # Title metadata
metadata_reso = ""  # Optional resolution metadata, e.g. 960,960
metadata_arch = ""  # Optional custom architecture metadata

# HuggingFace Settings
huggingface_repo_id = ""  # HuggingFace repository ID
huggingface_token = ""  # HuggingFace API token
huggingface_repo_type = ""  # Repository type
huggingface_repo_visibility = ""  # Repository visibility
huggingface_path_in_repo = ""  # Path within repository
save_state_to_huggingface = false  # Save training state to HuggingFace
resume_from_huggingface = ""  # Optional Hugging Face resume specification
async_upload = false  # Use async upload to HuggingFace