step2016/train_low.toml

# Wan 2.2 I2V 14B - LOW NOISE LoRA training
# Low noise model handles details, anatomy, fine textures

output_dir = '/workspace/work/output_low'
dataset = '/workspace/work/diffusion-pipe/configs/dataset_blink.toml'

# training settings
epochs = 1000
max_steps = 1875
micro_batch_size_per_gpu = 2
pipeline_stages = 1
gradient_accumulation_steps = 1
gradient_clipping = 1.0
warmup_steps = 100

# eval settings
eval_every_n_epochs = 1
eval_before_first_step = true
eval_micro_batch_size_per_gpu = 1
eval_gradient_accumulation_steps = 1

# misc settings
save_every_n_steps = 63
checkpoint_every_n_minutes = 60
activation_checkpointing = true
compile = true
partition_method = 'parameters'
save_dtype = 'bfloat16'
caching_batch_size = 1
steps_per_print = 1
video_clip_mode = 'single_beginning'

[model]
type = 'wan'
ckpt_path = '/workspace/work/models/Wan2.2-I2V-A14B'
transformer_path = '/workspace/work/models/comfyui/split_files/diffusion_models/wan2.2_i2v_low_noise_14B_fp16.safetensors'
llm_path = '/workspace/work/models/comfyui/split_files/text_encoders/umt5_xxl_fp16.safetensors'
dtype = 'bfloat16'
transformer_dtype = 'float8'
timestep_sample_method = 'logit_normal'
# Full timestep range for low noise — helps with high->low switching at inference
# (advice from PENISLORA author: train low on 0-1 instead of restricted range)

[adapter]
type = 'lora'
rank = 16
dtype = 'bfloat16'

[optimizer]
type = 'adamw_optimi'
lr = 2e-5
betas = [0.9, 0.99]
weight_decay = 0.01
eps = 1e-8