dataset.toml

resolutions = [ [ 1280, 704], [704, 1280]]
enable_ar_bucket = true
min_ar = 0.5
max_ar = 2.0
num_ar_buckets = 7
ar_buckets = [[1280, 704], [704, 1280]]
frame_buckets = [ 1, 48, 51, 72, 75, 81,]
[[directory]]
path = "/home/user/oiia_cat_videos_24fps"
num_repeats = 30

train.toml

output_dir = "/home/user/training_results/oiia_cat_videos_24fps/5B"
dataset =  "/home/user/dataset_5B.toml"
epochs = 20
micro_batch_size_per_gpu = 1
pipeline_stages = 1
gradient_accumulation_steps = 1
gradient_clipping = 1
warmup_steps = 100
eval_every_n_epochs = 1
eval_before_first_step = true
eval_micro_batch_size_per_gpu = 1
eval_gradient_accumulation_steps = 1
save_every_n_epochs = 1
activation_checkpointing = true
partition_method = "parameters"
save_dtype = "bfloat16"
caching_batch_size = 1
steps_per_print = 10
video_clip_mode = "single_beginning"
blocks_to_swap = 28

[model]
type = "wan"
ckpt_path = "/home/user/huggingface_wan/Wan2.2-TI2V-5B"
dtype = "bfloat16"
transformer_dtype = "float8"
timestep_sample_method = "logit_normal"

[adapter]
type = "lora"
rank = 64
dtype = "bfloat16"

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