# trained using diffusion-pipe commit 6e95020cad0b3cd7dcb93ce42b358669051bf6d2

#deepspeed --num_gpus=1 train.py --deepspeed --config /mnt/d/Programs/Neuro/diffusion-pipe/configs/main_anima_lora.toml --i_know_what_i_am_doing

#tensorboard --logdir /mnt/d/Programs/Neuro/diffusion-pipe/output_models/

output_dir = '/mnt/d/Programs/Neuro/diffusion-pipe/output_models'

dataset = '/mnt/d/Programs/Neuro/diffusion-pipe/datasets/anima_dataset.toml'

# training settings
epochs = 1000
max_steps = 600
micro_batch_size_per_gpu = 4
pipeline_stages = 1
gradient_accumulation_steps = 1
gradient_clipping = 1
warmup_steps = 50

# misc settings
save_every_n_epochs = 1
#save_every_n_steps = 100
activation_checkpointing = true

partition_method = 'parameters'

save_dtype = 'bfloat16'
caching_batch_size = 1
map_num_proc = 8
steps_per_print = 1
compile = true

[model]
type = 'anima'
transformer_path = '/mnt/d/Programs/Neuro/StabilityMatrix/Data/Models/DiffusionModels/anima_baseV10.safetensors'
vae_path = '/mnt/d/Programs/Neuro/StabilityMatrix/Data/Models/VAE/qwen_image_vae.safetensors'
llm_path = '/mnt/d/Programs/Neuro/StabilityMatrix/Data/Models/TextEncoders/qwen_3_06b_base.safetensors'
dtype = 'bfloat16'
#cache_text_embeddings = false
llm_adapter_lr = 0
sigmoid_scale = 1.3
timestep_sample_method = 'logit_normal'

[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