# Environment Configuration
launcher: "accelerate" # Options: accelerate
config_file: config/deepspeed/deepspeed_zero2.yaml # Path to distributed config file (optional)
num_processes: 16 # Number of processes to launch (overrides config file)
main_process_port: 29500
mixed_precision: "bf16" # Options: no, fp16, bf16
# Data Configuration
data:
datasets:
- name: tryon_synthetic_20k # Unique identifier (used in metrics, caching, reward routing)
dataset_dir: "/mnt/bn/image-data-lf/yunlong/data/TryOn/TryOn_Synthetic_NanoPro/version/0513_cqc20k" # Path to dataset folder
train: # Training participation config
weight: 1 # Mixing weight (integer); ratio with other sources determines batch allocation
max_dataset_size: 4096 # Cap on number of training samples for this source
eval: {} # Eval participation (inherits shared `eval:` section settings)
preprocessing_batch_size: 8 # Batch size for preprocessing
enable_preprocess: true # 启动离线预处理
dataloader_num_workers: 16 # Number of workers for DataLoader
force_reprocess: false # Force reprocessing of the dataset
cache_dir: "~/.cache/flow_factory/datasets" # Cache directory for preprocessed datasets
sampler_type: "auto" # Options: auto, distributed_k_repeat, group_contiguous
# Model Configuration
model:
finetune_type: 'lora' # Options: full, lora
lora_rank : 64
lora_alpha : 128
target_modules: "default" # Options: all, default, or list of module names like ["to_k", "to_q", "to_v", "to_out.0"]
model_name_or_path: "/mnt/bn/image-data-lf/yunlong/models/FLUX.2-klein-base-9B-RL" # Options: black-forest-labs/FLUX.2-klein-base-4B, black-forest-labs/FLUX.2-klein-base-9B
model_type: "flux2-klein-tryon"
resume_path: null # Path to load previous checkpoint/lora adapter
resume_type: null # Options: lora, full, state. Null to auto-detect based on `finetune_type`
log:
run_name: "flux2_lora_16gpus_nft_0604_nocfg_kl" # Run name (auto: {model_type}_{finetune_type}_{trainer_type}_{timestamp})
project: "Flow-Factory-Custom" # Project name for logging
logging_backend: "wandb" # Options: wandb, swanlab, none
save_dir: "/mnt/bn/image-data-lf/yunlong/Flow-Factory/train/NFT/" # Directory to save model checkpoints and logs
save_freq: 20 # Save frequency in epochs (0 to disable)
save_model_only: true # Save only the model weights (not optimizer, scheduler, etc.)
# Training Configuration
train:
# Trainer settings
trainer_type: 'nft'
advantage_aggregation: 'gdpo' # Options: 'sum', 'gdpo'
nft_beta: 0.1
# `Old` Policy settings
off_policy: true # Whether to use ema parameters for sampling off-policy data.
ema_decay_schedule: "piecewise_linear" # Decay schedule for EMA. Options: ['constant', 'power', 'linear', 'piecewise_linear', 'cosine', 'warmup_cosine']
flat_steps: 0
ramp_rate: 0.001
ema_decay: 0.5 # EMA decay rate (0 to disable)
ema_update_interval: 1 # EMA update interval (in epochs)
ema_device: "cuda" # Device to store EMA model (options: cpu, cuda)
# Training Timestep distribution
num_train_timesteps: 2 # Set null to all steps
time_sampling_strategy: discrete # Options: uniform, logit_normal, discrete, discrete_with_init, discrete_wo_init
time_shift: 3.0
timestep_range: 0.5 # Select fraction of timesteps to train on
# KL div
kl_type: 'v-based'
kl_beta: 1e-3 # KL divergence beta - 0 to disable and save memory. Set ~1e-2 for 'x-based' and ~1e-3 for 'v-based'.
ref_param_device: 'cuda' # Options: cpu, cuda
# Clipping
adv_clip_range: 5.0 # Advantage clipping range
# Sampling Settings
resolution: [1168, 880] # 训练时FM采样生成的分辨率。Can be int or [height, width]
condition_image_size: 1024 # 训练时FM输入编辑图像的分辨率。Can be int or [height, width]，注意实际用的时候是计算max_area
max_sequence_length: 1024 # 文本 Prompt 编码的最大截断长度 (默认512)
num_inference_steps: 8 # 训练时采样的时间步，对应的Flow-GRPO的Denoising Reduction。Number of timesteps
guidance_scale: 1.0 # Guidance scale for sampling
# Batch and sampling
per_device_batch_size: 1 # Batch size per device. For image-to-image task, this will always fallback to 1.
group_size: 16 # Group size for GRPO sampling
global_std: false # Use global std for advantage normalization
unique_sample_num_per_epoch: 48 # Unique samples per group
gradient_step_per_epoch: 1 # Gradient steps per epoch
gradient_accumulation_steps: auto # Options: auto, or positive integer. When set, `gradient_step_per_epoch` is ignored.
# Optimization
learning_rate: 1.0e-4 # Initial learning rate
adam_weight_decay: 1.0e-4 # AdamW weight decay
adam_betas: [0.9, 0.999] # AdamW betas
adam_epsilon: 1.0e-8 # AdamW epsilon
max_grad_norm: 1.0 # Max gradient norm for clipping
# Gradient Checkpointing
enable_gradient_checkpointing: true # Enable gradient checkpointing to save memory with extra compute
# offload_samples_to_cpu: CPU-offload sample tensor fields between sample() and optimize() to reduce GPU peak memory.
# Pros (true): saves N x per_batch_size GPU memory in sample()/optimize() (often GB-scale); no correctness or convergence impact.
# Cons (true): adds ~100ms/epoch H2D in reward path; tiny per-batch H2D in optimize (<5ms each).
# Recommended (true) for higher resolutions, larger batch sizes, or any sample()/optimize() OOM. Default false works for current example settings.
offload_samples_to_cpu: false
# Seed
seed: 0 # Random seed
# Scheduler Configuration
scheduler:
dynamics_type: "ODE" # Options: Flow-SDE, Dance-SDE, CPS, ODE
# Evaluation settings
eval:
resolution: [1168, 880] # Evaluation resolution
condition_image_size: 1024
max_sequence_length: 1024 # 文本 Prompt 编码的最大截断长度 (默认512)
per_device_batch_size: 1 # Eval batch size
guidance_scale: 1.0 # Guidance scale for sampling
num_inference_steps: 28 # Number of eval timesteps
eval_freq: 20 # Eval frequency in epochs (0 to disable)
seed: 0 # Eval seed (defaults to training seed)
# Reward Model Configuration
rewards:
- name: "face_consistency"
reward_model: "facezhichuangreward"
weight: 1.0
batch_size: 16
timeout: 10.0
max_area: 1048576 # 1024 * 1024
person_image_index: 1 # TryOn 任务中，condition_images[1] 通常是人像图
max_concurrent: 16 # RPC 并发请求限制
async_reward: true # 开启异步并发打分，隐藏网络延迟
「项目背景」输入服饰+人脸，输出TryOn结果
「训练config文件」
「补充说明」其中奖励函数是人脸相似度
「训练曲线」
「问题」