chroma-kwis-v1 / config.toml
zaksynack's picture
Added PEFT and diffusion-pipe config data
4ec3389 verified
# Dataset config file.
output_dir = '/root/outputs'
dataset = 'configs/dataset.toml'
# Training settings
epochs = 50
micro_batch_size_per_gpu = 5
pipeline_stages = 1
gradient_accumulation_steps = 1
gradient_clipping = 1.0
warmup_steps = 20
# eval settings
eval_every_n_epochs = 1
eval_before_first_step = false
eval_micro_batch_size_per_gpu = 5
eval_gradient_accumulation_steps = 1
# misc settings
save_every_n_epochs = 2
checkpoint_every_n_epochs = 2
activation_checkpointing = true
partition_method = 'parameters'
save_dtype = 'bfloat16'
caching_batch_size = 3
steps_per_print = 1
compile = true
video_clip_mode = 'single_beginning'
#blocks_to_swap = 15
[model]
type = 'wan'
ckpt_path = '/root/outputs/models/Wan2.1-T2V-14B'
#diffusers_path = '/root/outputs/models/FLUX.1-dev'
#transformer_path = '/root/outputs/models/chroma/Chroma.safetensors'
llm_path = '/root/outputs/models/Wan2.1-T2V-14B/models_t5_umt5-xxl-enc-bf16.pth'
dtype = 'bfloat16'
# You can optionally load the transformer in fp8 when training LoRAs.
transformer_dtype = 'float8'
timestep_sample_method = 'logit_normal'
#flux_shift = true
[adapter]
type = 'lora'
rank = 32
dtype = 'bfloat16'
#init_from_existing = '/root/outputs/5c6d31124f144544913effcc0a17e0ea/epoch10'
[optimizer]
type = 'adamw_optimi'
lr = 2e-4
betas = [0.9, 0.99]
weight_decay = 0.01
eps = 1e-8
[monitoring]
enable_wandb = true
wandb_api_key = '316edc68e8c9d21674a9bb887bc5c599e2c94e35'
wandb_tracker_name = 'wan'
wandb_run_name = 'wan-kwis-v1'