checkpoints/step2000/train.toml

# KREA2 edit saga 2026-07-29 — treino principal.
#
# Base: receita validada krea2_multiref_grounded (docs/KREA2_MULTIREF_RECEITA.md),
# aqui com max_refs=1 (todos os datasets são pares A->B).
# - width_shift + refs a t=0 + LoRA condition-only routada + grounding Qwen3-VL
# - caption_dropout=0.1 (uncond do CFG; condition_dropout é proibido no grounded)
# - txtfusion_rank=128 (canal de leitura reforçado — tarefa é semântica: editar)
# - rank 64 (usuário pediu >32; dezenas de milhares de imagens)

output_dir = '/workspace/checkpoints/krea2_edit_saga'
dataset = 'examples/krea2_edit_saga/dataset.toml'
epochs = 1000
max_steps = 22500            # 3 épocas × 30.000 pares / batch efetivo 4
micro_batch_size_per_gpu = 1
gradient_accumulation_steps = 4
pipeline_stages = 1
gradient_clipping = 1.0
warmup_steps = 100
save_every_n_steps = 500
checkpoint_every_n_minutes = 120
activation_checkpointing = true
save_dtype = 'bfloat16'
caching_batch_size = 1       # OBRIGATÓRIO com multi_ref
blocks_to_swap = 0           # usuário pediu velocidade; VRAM medida 20,1G com swap 8, cabe sem

[model]
type = 'krea2_multiref_grounded'
diffusion_model = '/workspace/models/krea2/diffusion_models/krea2_raw_fp8_scaled.safetensors'
vae = '/workspace/models/krea2/split_files/vae/qwen_image_vae.safetensors'
text_encoders = [
    {path = '/workspace/models/krea2/text_encoders/qwen3vl_4b_bf16.safetensors', type = 'krea2'},
]
dtype = 'bfloat16'
diffusion_model_dtype = 'float8'
timestep_sample_method = 'logit_normal'
flux_shift = true

[krea2_multiref_grounded]
max_refs = 1
slot_axis = 'width'
position_mode = 'width_shift'
reference_position_offset = 1.0
reference_timestep = 'zero'
condition_dropout = 0.0      # proibido no caminho grounded
condition_only_lora = true
vl_prompt_style = 'picture_n'
vl_image_label = 'image'
vl_image_max_pixels = 147456 # 384² — grounding; cache é o gargalo, medir no smoke
caption_dropout = 0.1
txtfusion_rank = 128

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

[optimizer]
type = 'AdamW8bitKahan'
lr = 1e-4
betas = [0.9, 0.99]
weight_decay = 0.01
stabilize = false